Chapter6.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Chapter 6: Graphs\n",
    "\n",
    "Accompanying code for the book [The Art of Feature Engineering](http://artoffeatureengineering.com).\n",
    "\n",
    "This notebook plus notebooks for the other chapters are available online at https://github.com/DrDub/artfeateng\n",
    "\n",
    "\n",
    "## MIT License\n",
    "\n",
    "Copyright 2019 Pablo Duboue\n",
    "\n",
    "Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the \"Software\"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:\n",
    "\n",
    "The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.\n",
    "\n",
    "THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Limitations\n",
    "\n",
    "* Simple python intented for people coming from other languages that plan to use the ideas described in the book outside of Python.\n",
    "* Many of these techniques are available as library calls. They are spelled out as for teaching purposes.\n",
    "* Resource limitations:\n",
    " * At most one day of running time per notebook.\n",
    " * No GPU required.\n",
    " * Minimal dependencies.\n",
    " * At most 8Gb of RAM.\n",
    "* Due to resource limitations, these notebooks do not undergo as much hyperparameter tuning as necessary. This is a shortcoming of these case studies, keep it in mind if you want to follow a similar path with your experiments.\n",
    "* To help readers try variants of some cells in isolation, the cells are easily executable without having to re-run the whole notebook. As such, most cells read everything they need from disk and write all their results back into disk, which is unnecessary with normal notebooks. The code for each cell might look long and somewhat unusual. In a sense, each cell tries to be a separate Python program.\n",
    "* I dislike Pandas so these notebooks are Pandas-free, which might seem unusual to some."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Chapter 6: Case Study on Graph Data\n",
    "\n",
    "In this chapter I will describe the creation of the base WikiCities dataset employed in this chapter and the next three chapters.\n",
    "\n",
    "## WikiCities Dataset\n",
    "\n",
    "It starts with the latest release of DBpedia [cite], for October 2016, a curated ontological resource obtained from the infoboxes at Wikipedia. Infoboxes are the tabular information appearing next to the page content in each Wikipedia page.\n",
    "\n",
    "My intention with this dataset is to provide a task that can be attacked with regular features, with time series features, textual features and image features. The task I found that allow for all these approaches is that of calculating the population of a city or town based on their ontological properties (e.g., name of its leader or its time zone), based on its historical population and historical features (which involves a time series analysis), based on the textual description of the place (which involves text analysis, particularly as many times further includes the population) and a satellite image of the city (which involves image processing).\n",
    "\n",
    "DBpedia files are distributed as files of triples, where a triple consists of a subject, verb and object. The subject is always an entity in the ontology, prefixed by the name space dbpedia.org. The verb is an ontological relation, of which there are many (type, leader-name, etc). The object can be either another entity or a literal value (a string or a number). Besides the actual files, DBpedia also contains links to multiple other ontologies, like Freebase or Yago plus links to the original Wikipedia articles that resulted in the extracted information. For these case studies we will be using the \"instance_types_en.ttl.bz2\", \"mappingbased_literals_en.ttl.bz2\", \"mappingbased_objects_en.ttl.bz2\".\n",
    "\n",
    "The type information is derived from the infobox itself. The first challenge is to identify cities, as the \"city\" or \"settlement\" type is not thoroughly annotated in Wikipedia. Instead of relying on heuristics (such as \"any entitity with a location and a population\" which will result in neighborghoods annotated), I chose to use an external more reliable source, the GeoNames project. The project distributes a list of \"cities or settlements with at least 1000 people\". It contains 128,000+ places (file \"cities1000.txt\"). To link them back to DBpedia we use the file \"geonames_links.ttl.bz2\" distributed by DBpedia (Cell #1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Cities1000 total: 128,764\n",
      "In DBpedia found: 80,251\n"
     ]
    }
   ],
   "source": [
    "# CELL 1\n",
    "import re\n",
    "import bz2\n",
    "\n",
    "id_re = re.compile(\"^(\\d+)\")\n",
    "url_re = re.compile(\".*geonames\\.org\\/(\\d+)\")\n",
    "\n",
    "cities1000_ids = set()\n",
    "with open(\"cities1000.txt\") as cities_file:\n",
    "    for line in cities_file:\n",
    "        cities1000_ids.add(id_re.match(line).group(0))\n",
    "print(\"Cities1000 total: {:,}\".format(len(cities1000_ids)))\n",
    "\n",
    "found = 0\n",
    "with open(\"ch6_cell1_cities1000_links.ttl\",\"w\") as out:\n",
    "    with bz2.BZ2File(\"geonames_links.ttl.bz2\",\"r\") as links:\n",
    "        for byteline in links:\n",
    "            if byteline[0] == ord('#'):\n",
    "                continue\n",
    "            line = byteline.decode(\"utf-8\")\n",
    "            if not url_re.match(line):\n",
    "                print(byteline)\n",
    "            if url_re.match(line).group(1) in cities1000_ids:\n",
    "                out.write(line)\n",
    "                found += 1            \n",
    "print(\"In DBpedia found: {:,}\".format(found))\n",
    "cities1000_ids = None # free memory"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Therefore, using the available links to DBpedia, we found 80,251 of them in the Wikipedia data set. These 80,251 plus DBpedia consitute the raw data for this problem (files \"ch6_cell1_cities1000_links.ttl\", ). \n",
    "\n",
    "Following the methodology described in Chapter 1, I will set aside 20% of this data for final evaluation at the end of the feature engineering process (Cell #2)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Devset size: 64,200\n",
      "Heldout size: 16,051\n"
     ]
    }
   ],
   "source": [
    "# CELL 2\n",
    "import random\n",
    "\n",
    "rand = random.Random(42)\n",
    "all_cities = open(\"ch6_cell1_cities1000_links.ttl\").readlines()\n",
    "rand.shuffle(all_cities)\n",
    "pivot = int(len(all_cities) * 0.8)\n",
    "devset = all_cities[:pivot]\n",
    "heldout = all_cities[pivot:]\n",
    "with open(\"ch6_cell2_cities1000_dev_links.ttl\", \"w\") as dev:\n",
    "    dev.write(\"\".join(devset))\n",
    "with open(\"ch6_cell2_cities1000_held_links.ttl\", \"w\") as held:\n",
    "    held.write(\"\".join(heldout))\n",
    "print(\"Devset size: {:,}\".format(len(devset)))\n",
    "print(\"Heldout size: {:,}\".format(len(heldout)))\n",
    "all_cities = devset = heldout = None # free memory"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Following the methodology in Chapter 1, I now want to do some exploratory data analysis over these 64,200 settlements to decide basic featurization and model to employ.\n",
    "\n",
    "The key issue to resolve here is to drill down on the specific relations we think will be useful to predict our target class. While DBpedia contain thousands of relations, many are seldom expressed (meaning, there are few pages where a human editor feel the need to record information such as \"distante-to-London\").\n",
    "\n",
    "While we want to account the number of relations where our cities of interest partake, something to remember is that they might be value in relations that have a given city as its subject but there might also be value in relations that have the city as its object (that is, the inverse of certain relations).\n",
    "\n",
    "Because processing the large files of DB is a computationally onerous process, I'll start by pre-filtering the triples related to all entities both forward and inverse, generating the file \"ch6_cell3_cities1000_base.ttl\" (Cell #3)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Base 2,014,559 out of 38,285,143\n"
     ]
    }
   ],
   "source": [
    "# CELL 3\n",
    "city_uris = set(map(lambda line:line.split(' ')[0], open(\"ch6_cell1_cities1000_links.ttl\").readlines()))\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "def compressed_triples_gen(filename):\n",
    "    with bz2.BZ2File(filename,\"r\") as compressed:\n",
    "        for byteline in compressed:\n",
    "            if byteline[0] == ord('#'):\n",
    "                continue\n",
    "            line = byteline.decode(\"utf-8\")\n",
    "            s,v,o = triple_re.match(line).groups()\n",
    "            yield line,s,v,o\n",
    "            \n",
    "written = 0\n",
    "read = 0\n",
    "with open(\"ch6_cell3_cities1000_base.ttl\",\"w\") as base:\n",
    "    # types and literals only in subject position\n",
    "    for line, subj, verb, obj in compressed_triples_gen(\"instance_types_en.ttl.bz2\"):\n",
    "        read += 1 \n",
    "        if subj in city_uris:\n",
    "            base.write(line)\n",
    "            written += 1\n",
    "    for line, subj, verb, obj in compressed_triples_gen(\"mappingbased_literals_en.ttl.bz2\"):\n",
    "        read += 1\n",
    "        if subj in city_uris:\n",
    "            base.write(line)\n",
    "            written += 1\n",
    "    for line, subj, verb, obj in compressed_triples_gen(\"mappingbased_objects_en.ttl.bz2\"):\n",
    "        read += 1\n",
    "        if subj in city_uris:\n",
    "            base.write(line)\n",
    "            written += 1\n",
    "        if obj in city_uris:\n",
    "            base.write(\"{} {} {} .\\n\".format(obj, verb[0:-1] + \"?inv>\", subj))\n",
    "            written += 1\n",
    "print(\"Base {:,} out of {:,}\".format(written, read))\n",
    "city_uris = None # free memory"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "That code takes a while to run but it will speed up the rest of the work quite a bit down the road as we only need to worry now of 2 million triples instead of 38 million. The next step is to produce a filtered version just for the development and heldout sets into the files \"ch6_cell4_cities1000_devset_base.ttl\" and \"ch6_cell4_cities1000_heldout_base.ttl\", respectively (Cell #4)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Devset base size: 1,625,988 triples\n"
     ]
    }
   ],
   "source": [
    "# CELL 4\n",
    "devset_uris = set(map(lambda line:line.split(' ')[0], open(\"ch6_cell2_cities1000_dev_links.ttl\").readlines()))\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "devset_size = 0\n",
    "with open(\"ch6_cell3_cities1000_base.ttl\") as base:\n",
    "    with open(\"ch6_cell4_cities1000_dev_base.ttl\", \"w\") as devset:\n",
    "        with open(\"ch6_cell4_cities1000_heldout_base.ttl\", \"w\") as heldout:\n",
    "            for line in base:\n",
    "                subj, verb, obj = triple_re.match(line).groups()\n",
    "                if subj in devset_uris:\n",
    "                    devset.write(line)\n",
    "                    devset_size += 1\n",
    "                else:\n",
    "                    heldout.write(line)\n",
    "print(\"Devset base size: {:,} triples\".format(devset_size))\n",
    "devset_uris = None # free memory"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "These close to 1.6 million triples are our working set, \n",
    "\n",
    "## EDA: Exploratory Data Analysis\n",
    "\n",
    "I'll start the Exploratory Data Analysis by visualizing 10 random entities using the Graphviz package (Cell #5). The cities are rendered as ch6_cell5_city_*.pdf"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Rendering Doolittle,_Texas : 4 dot lines\n",
      "Rendering Belton,_Missouri : 62 dot lines\n",
      "Rendering Poiana_Sibiului : 37 dot lines\n",
      "Rendering Unadilla,_Georgia : 54 dot lines\n",
      "Rendering Sankeshwar,_Karnataka : 4 dot lines\n",
      "Rendering Psevdas : 22 dot lines\n",
      "Rendering Skudai : 41 dot lines\n",
      "Rendering Santa_Maria_da_Boa_Vista : 4 dot lines\n",
      "Rendering Newstead,_Tasmania : 18 dot lines\n",
      "Rendering Gombi : 20 dot lines\n",
      "\n",
      "Top relation counts:\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<table><tr><th>Rel</th><th>Counts</th></tr><tr><td>utcOffset</td><td>10</td></tr>\n",
       "<tr><td>isPartOf</td><td>10</td></tr>\n",
       "<tr><td>timeZone</td><td>8</td></tr>\n",
       "<tr><td>22-rdf-syntax-ns#type</td><td>7</td></tr>\n",
       "<tr><td>name</td><td>7</td></tr>\n",
       "<tr><td>country</td><td>7</td></tr>\n",
       "<tr><td>populationTotal</td><td>5</td></tr>\n",
       "<tr><td>areaTotal</td><td>4</td></tr>\n",
       "<tr><td>leaderTitle</td><td>4</td></tr>\n",
       "<tr><td>elevation</td><td>4</td></tr></table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 5\n",
    "import re\n",
    "import graphviz\n",
    "import random\n",
    "\n",
    "PARAM_USE_SET_FOR_BOOK = True\n",
    "\n",
    "devset_uris = set(map(lambda line:line.split(' ')[0], open(\"ch6_cell1_cities1000_links.ttl\").readlines()))\n",
    "rand = random.Random(42)\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "# stable set for book\n",
    "cities = [ \"Doolittle,_Texas\", \"Belton,_Missouri\", \"Poiana_Sibiului\", \"Unadilla,_Georgia\",\n",
    "          \"Sankeshwar,_Karnataka\", \"Psevdas\", \"Skudai\", \"Santa_Maria_da_Boa_Vista\",\n",
    "          \"Newstead,_Tasmania\", \"Gombi\" ]\n",
    "\n",
    "if PARAM_USE_SET_FOR_BOOK:\n",
    "    cities = list(map(lambda x: \"<http://dbpedia.org/resource/\" + x + \">\", cities))\n",
    "else:\n",
    "    cities = set(rand.sample(sorted(devset_uris), 10))\n",
    "\n",
    "nodes = dict()\n",
    "dots = dict()\n",
    "\n",
    "for city_uri in cities:\n",
    "    dots[city_uri] = graphviz.Digraph(format='pdf', graph_attr={'rankdir':'BT'})\n",
    "    nodes[city_uri] = dict()\n",
    "\n",
    "def cell5_make_or_return_node(city_uri, name, shape=None):\n",
    "    if name in nodes[city_uri]:\n",
    "        return nodes[city_uri][name]\n",
    "    ret = 'N' + str(len(nodes[city_uri]))\n",
    "    nodes[city_uri][name] = ret\n",
    "    if shape is None:\n",
    "        dots[city_uri].node(ret, name)\n",
    "    else:\n",
    "        dots[city_uri].node(ret, name.replace('&', ' '), shape=shape)\n",
    "    return ret\n",
    "\n",
    "for city_uri in cities:\n",
    "    cell5_make_or_return_node(city_uri, city_uri, 'star')\n",
    "    \n",
    "rel_count = dict()\n",
    "with open(\"ch6_cell4_cities1000_dev_base.ttl\") as base:\n",
    "    for line in base:\n",
    "        subj, verb, obj = triple_re.match(line).groups()\n",
    "        if subj in cities:\n",
    "            city_uri = subj\n",
    "            rel_count[verb] = rel_count.get(verb, 0) + 1\n",
    "            if verb.endswith(\"?inv>\"):\n",
    "                subj, obj = obj, subj\n",
    "                verb = verb[:-5] + \">\"\n",
    "            subj_name = cell5_make_or_return_node(city_uri, subj)\n",
    "            obj_name = cell5_make_or_return_node(city_uri, obj)\n",
    "            dots[city_uri].edge(subj_name, obj_name, verb)\n",
    "for city_uri in cities:\n",
    "    city = city_uri[:-1].split(\"/\")[-1]\n",
    "    dot = dots[city_uri]\n",
    "    print(\"Rendering\",city,\":\",len(dot.source.splitlines()), \"dot lines\")\n",
    "    dot.render(\"ch6_cell5_city_\" + city + \".dot\")\n",
    "with open(\"ch6_cell5_rel_counts.tsv\", \"w\") as rels:\n",
    "    for rel in sorted(rel_count.keys()):\n",
    "        rels.write(\"{}\\t{}\\n\".format(rel,rel_count[rel]))\n",
    "print(\"\\nTop relation counts:\")\n",
    "from IPython.display import HTML, display\n",
    "display(HTML(\"<table><tr><th>Rel</th><th>Counts</th></tr>\" +\n",
    "            \"\\n\".join(list(map(lambda r:\"<tr><td>{}</td><td>{}</td></tr>\".format(r[0][:-1].split(\"/\")[-1],r[1]), \n",
    "                               sorted(rel_count.items(), key=lambda t:t[1], reverse=True)[:10]))) +\n",
    "            \"</table>\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "From here we can see that some randomly sampled cities are all fairly small, which talks that most rows will have very little data while others will contain the bulk of the triples. The amount of infrormation available is already a great indicator of population.\n",
    "\n",
    "From looking at the graphs we can see that certain places (like Skudai) do not contain population information. The next step is thus to filter cities for which the population is known (as we can only use those to build a regressor or classifier).\n",
    "\n",
    "From the relation table in \"ch6_cell5_rel_counts.tsv\" we can see that some relations are very rare (like homepage or leaderParty) while others are defined for almost all cities (like rdf type or areaLand). The inverse relations are also quite sparse but these places are all sparsely populated.\n",
    "\n",
    "But first we will need to know how population is expressed in DBpedia. From Psevdas we can see it has a relation \"populationTotal\". Would that be the only way population is expressed? Let us see what relations contain the word population in their name (Cell #6)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<table><tr><th>Rel</th><th>Counts</th></tr><tr><td>populationTotal</td><td>49028</td></tr>\n",
       "<tr><td>populationDensity</td><td>19540</td></tr>\n",
       "<tr><td>populationAsOf</td><td>5410</td></tr>\n",
       "<tr><td>populationPlace?inv</td><td>2511</td></tr>\n",
       "<tr><td>populationMetro</td><td>1286</td></tr>\n",
       "<tr><td>populationUrban</td><td>1188</td></tr>\n",
       "<tr><td>populationTotalRanking</td><td>1057</td></tr>\n",
       "<tr><td>populationMetroDensity</td><td>142</td></tr>\n",
       "<tr><td>populationUrbanDensity</td><td>79</td></tr>\n",
       "<tr><td>populationRural</td><td>42</td></tr></table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 6\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "rels_of_interest = dict()\n",
    "with open(\"ch6_cell4_cities1000_dev_base.ttl\") as base:\n",
    "    for line in base:\n",
    "        subj, verb, obj = triple_re.match(line).groups()\n",
    "        if \"population\" in verb:\n",
    "            rels_of_interest[verb] = rels_of_interest.get(verb, 0) + 1\n",
    "from IPython.display import HTML, display\n",
    "display(HTML(\"<table><tr><th>Rel</th><th>Counts</th></tr>\" +\n",
    "            \"\\n\".join(list(map(lambda r:\"<tr><td>{}</td><td>{}</td></tr>\".format(r[0][:-1].split(\"/\")[-1],r[1]), \n",
    "                               sorted(rels_of_interest.items(), key=lambda t:t[1], reverse=True)))) +\n",
    "            \"</table>\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "It seems \"populationTotal\" is the right relation to use, but we will need to filter populationMetro, populationUrban and populationRural. Also, if populationTotal is not defined, either of those, if defined could be used as surrogates of out target class. But at this stage we can to filter the cities with defined population, I will put them into the file \"ch6_cell7_dev_pop.tsv\" and collect also the population (will require cleaning type information such as ^^<http://www.w3.org/2001/XMLSchema#nonNegativeInteger>). This number will be our target value for regression (or a discretized version of it, our target class).\n",
    "\n",
    "I do this process on both the dev and heldout (into file \"ch6_cell7_heldout_pop.tsv\") in Cell #7."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Cities with known population in devset: 48,868\n"
     ]
    }
   ],
   "source": [
    "# CELL 7\n",
    "import re\n",
    "from collections import OrderedDict\n",
    "\n",
    "cities_uris = set(map(lambda line:line.split(' ')[0], open(\"ch6_cell1_cities1000_links.ttl\").readlines()))\n",
    "devset_uris = set(map(lambda line:line.split(' ')[0], open(\"ch6_cell2_cities1000_dev_links.ttl\").readlines()))\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "population = OrderedDict()\n",
    "with open(\"ch6_cell3_cities1000_base.ttl\") as base:\n",
    "    for line in base:\n",
    "        subj, verb, obj = triple_re.match(line).groups()\n",
    "        verb = verb[:-1].split(\"/\")[-1]\n",
    "        update = False\n",
    "        conditional = True\n",
    "        if verb == \"populationTotal\":\n",
    "            update = True\n",
    "            conditional = False\n",
    "        elif verb in set(['populationMetro', 'populationUrban', 'populationRural']):\n",
    "            if subj in population and population[subj][1] == 'populationTotal':\n",
    "                pass # keep current\n",
    "            else:\n",
    "                update = True\n",
    "                conditional = subj in population\n",
    "        if update:\n",
    "            # compute actual population\n",
    "            if obj[0] == '\"':\n",
    "                obj = obj[1:obj[1:].index('\"')+1]\n",
    "            try:\n",
    "                pop = float(obj)\n",
    "            except ValueError:\n",
    "                print(\"For city\", subj, \"cannot convert\", line)\n",
    "                next # ignore\n",
    "            if not conditional or population[subj][0] < pop:\n",
    "                population[subj] = (pop, verb)\n",
    "                \n",
    "dev_cities_with_pop = 0\n",
    "with open(\"ch6_cell7_dev_pop.tsv\", \"w\") as dev:\n",
    "    with open(\"ch6_cell7_heldout_pop.tsv\", \"w\") as held:\n",
    "        for city, pop in population.items():\n",
    "            line = \"{}\\t{}\\n\".format(city, pop[0])\n",
    "            if city in devset_uris:\n",
    "                dev.write(line)\n",
    "                dev_cities_with_pop += 1\n",
    "            else:\n",
    "                held.write(line)\n",
    "print(\"Cities with known population in devset: {:,}\".format(dev_cities_with_pop))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now we can proceed to repeat the plotting of 10 random cities in the devset with known population (Cell #8)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Rendering Ovalle : 59 dot lines\n",
      "Rendering Hawaiian_Paradise_Park,_Hawaii : 48 dot lines\n",
      "Rendering Coupar_Angus : 4 dot lines\n",
      "Rendering Hilsenheim : 28 dot lines\n",
      "Rendering Esko,_Minnesota : 37 dot lines\n",
      "Rendering Hamlin,_New_York : 46 dot lines\n",
      "Rendering Falkenberg,_Märkisch-Oderland : 32 dot lines\n",
      "Rendering Wermelskirchen : 45 dot lines\n",
      "Rendering Berne,_Indiana : 67 dot lines\n",
      "Rendering Chennevières-sur-Marne : 20 dot lines\n",
      "\n",
      "Top relation counts:\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<table><tr><th>Rel</th><th>Counts</th></tr><tr><td>isPartOf</td><td>11</td></tr>\n",
       "<tr><td>name</td><td>10</td></tr>\n",
       "<tr><td>areaCode</td><td>10</td></tr>\n",
       "<tr><td>22-rdf-syntax-ns#type</td><td>9</td></tr>\n",
       "<tr><td>elevation</td><td>9</td></tr>\n",
       "<tr><td>populationTotal</td><td>9</td></tr>\n",
       "<tr><td>country</td><td>9</td></tr>\n",
       "<tr><td>areaTotal</td><td>8</td></tr>\n",
       "<tr><td>postalCode</td><td>8</td></tr>\n",
       "<tr><td>birthPlace?inv</td><td>8</td></tr></table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 8\n",
    "import re\n",
    "import graphviz\n",
    "import random\n",
    "\n",
    "PARAM_USE_SET_FOR_BOOK = True\n",
    "\n",
    "devset_uris = set(map(lambda line:line.split('\\t')[0], open(\"ch6_cell7_dev_pop.tsv\").readlines()))\n",
    "rand = random.Random(42)\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "# stable set for book\n",
    "cities = [ \"Ovalle\", \"Hawaiian_Paradise_Park,_Hawaii\", \"Coupar_Angus\",\n",
    "          \"Hilsenheim\", \"Esko,_Minnesota\", \"Hamlin,_New_York\", \"Falkenberg,_Märkisch-Oderland\",\n",
    "          \"Wermelskirchen\", \"Berne,_Indiana\", \"Chennevières-sur-Marne\" ]\n",
    "\n",
    "if PARAM_USE_SET_FOR_BOOK:\n",
    "    cities = list(map(lambda x: \"<http://dbpedia.org/resource/\" + x + \">\", cities))\n",
    "else:\n",
    "    cities = set(rand.sample(sorted(devset_uris), 10))\n",
    "\n",
    "nodes = dict()\n",
    "dots = dict()\n",
    "\n",
    "for city_uri in cities:\n",
    "    dots[city_uri] = graphviz.Digraph(format='pdf', graph_attr={'rankdir':'BT'})\n",
    "    nodes[city_uri] = dict()\n",
    "\n",
    "for city_uri in cities:\n",
    "    cell5_make_or_return_node(city_uri, city_uri, 'star')\n",
    "\n",
    "rel_count = dict()\n",
    "with open(\"ch6_cell4_cities1000_dev_base.ttl\") as base:\n",
    "    for line in base:\n",
    "        subj, verb, obj = triple_re.match(line).groups()\n",
    "        if subj in cities:\n",
    "            city_uri = subj\n",
    "            rel_count[verb] = rel_count.get(verb, 0) + 1\n",
    "            if verb.endswith(\"?inv>\"):\n",
    "                subj, obj = obj, subj\n",
    "                verb = verb[:-5] + \">\"\n",
    "            subj_name = cell5_make_or_return_node(city_uri, subj)\n",
    "            obj_name = cell5_make_or_return_node(city_uri, obj)\n",
    "            dots[city_uri].edge(subj_name, obj_name, verb)\n",
    "for city_uri in cities:\n",
    "    city = city_uri[:-1].split(\"/\")[-1]\n",
    "    dot = dots[city_uri]\n",
    "    print(\"Rendering\",city,\":\",len(dot.source.splitlines()), \"dot lines\")\n",
    "    dot.render(\"ch6_cell8_city_\" + city + \".dot\")\n",
    "    \n",
    "with open(\"ch6_cell8_rel_counts.tsv\", \"w\") as rels:\n",
    "    for rel in sorted(rel_count.keys()):\n",
    "        rels.write(\"{}\\t{}\\n\".format(rel,rel_count[rel]))\n",
    "print(\"\\nTop relation counts:\")\n",
    "from IPython.display import HTML, display\n",
    "display(HTML(\"<table><tr><th>Rel</th><th>Counts</th></tr>\" +\n",
    "            \"\\n\".join(list(map(lambda r:\"<tr><td>{}</td><td>{}</td></tr>\".format(r[0][:-1].split(\"/\")[-1],r[1]), \n",
    "                               sorted(rel_count.items(), key=lambda t:t[1], reverse=True)[:10]))) +\n",
    "            \"</table>\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "These cities with defined population have definitely much more things going on for them, including multiple areaCode relations (three in the case of Wermelskirchen) and a variety of inverse relations. At this stage on the EDA process we want to look into which of these relations might make for good features. Going back to Chapter 1, we want features that will be simple, related to the target (population) and readily available. From the perspective of \"readily available\" we want to shorten down the number of relations to ones that are frequent enough that will be useful for a ML algorithm (I will look into the other issues later). Therefore their frequency over the whole set of entities is paramount. Let us sort the relations by the number of cities that have them defined (Cell #9). "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Total relations 347 percentile size 34\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<table><tr><th>Pos</th><th>Rel</th><th>Counts</th><th>Percentage of cities defined</th></tr><tr><td>First</td><td>buildingEndDate</td><td>1</td><td>0.0%</td></tr>\n",
       "<tr><td>At 10 percentile</td><td>sportCountry?inv</td><td>1</td><td>0.0%</td></tr>\n",
       "<tr><td>At 20 percentile</td><td>lowestState</td><td>3</td><td>0.0%</td></tr>\n",
       "<tr><td>At 30 percentile</td><td>inflow?inv</td><td>5</td><td>0.01%</td></tr>\n",
       "<tr><td>At 40 percentile</td><td>board?inv</td><td>9</td><td>0.01%</td></tr>\n",
       "<tr><td>At 50 percentile</td><td>riverMouth?inv</td><td>27</td><td>0.05%</td></tr>\n",
       "<tr><td>At 60 percentile</td><td>county?inv</td><td>78</td><td>0.15%</td></tr>\n",
       "<tr><td>At 70 percentile</td><td>knownFor?inv</td><td>209</td><td>0.42%</td></tr>\n",
       "<tr><td>At 80 percentile</td><td>populationTotalRanking</td><td>699</td><td>1.43%</td></tr>\n",
       "<tr><td>At 90 percentile</td><td>headquarter?inv</td><td>2769</td><td>5.66%</td></tr>\n",
       "<tr><td>At 100 percentile</td><td>areaTotal</td><td>29917</td><td>61.22%</td></tr>\n",
       "<tr><td>Last</td><td>22-rdf-syntax-ns#type</td><td>48868</td><td>100.0%</td></tr></table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 9\n",
    "import re\n",
    "from collections import OrderedDict\n",
    "\n",
    "devset_uris = set(map(lambda line:line.split('\\t')[0], open(\"ch6_cell7_dev_pop.tsv\").readlines()))\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "rel_count = OrderedDict()\n",
    "rel_subj_seen = set()\n",
    "with open(\"ch6_cell4_cities1000_dev_base.ttl\") as base:\n",
    "    for line in base:\n",
    "        subj, verb, obj = triple_re.match(line).groups()\n",
    "        if subj in devset_uris:\n",
    "            rel_subj = \"{}-{}\".format(subj,verb)\n",
    "            if rel_subj not in rel_subj_seen:\n",
    "                rel_count[verb] = rel_count.get(verb, 0) + 1\n",
    "                rel_subj_seen.add(rel_subj)\n",
    "sorted_count = sorted(rel_count.items(), key=lambda t:t[1])\n",
    "\n",
    "# drop long names\n",
    "sorted_count = list(map(lambda t:(t[0][:-1].split(\"/\")[-1], str(t[1])), sorted_count))\n",
    "table = list()\n",
    "step = int(len(sorted_count) / 10)\n",
    "print(\"Total relations\", len(sorted_count), \"percentile size\", step)\n",
    "table.append( (\"First\", sorted_count[0][0], sorted_count[0][1], \n",
    "              \"{}%\".format(int(float(sorted_count[0][1]) / len(devset_uris) * 10000) / 100) ))\n",
    "current = step\n",
    "percentile = 1\n",
    "while current < len(sorted_count):\n",
    "    table.append( (\"At {} percentile\".format(percentile*10), sorted_count[current][0],\n",
    "                 sorted_count[current][1],\n",
    "                 \"{}%\".format(int(float(sorted_count[current][1]) / len(devset_uris) * 10000) / 100) ) )\n",
    "    current += step\n",
    "    percentile += 1\n",
    "table.append( (\"Last\", sorted_count[-1][0], sorted_count[-1][1], \n",
    "              \"{}%\".format(int(float(sorted_count[-1][1]) / len(devset_uris) * 10000) / 100) ))\n",
    "\n",
    "\n",
    "from IPython.display import HTML, display\n",
    "display(HTML(\"<table><tr><th>Pos</th><th>Rel</th><th>Counts</th><th>Percentage of cities defined</th></tr>\" +\n",
    "            \"\\n\".join(list(map(lambda r:\"<tr><td>\" + \"</td><td>\".join(r) + \"</td></tr>\", table))) +\n",
    "            \"</table>\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Which relation we will use is key to success of the ML, at this stage we will overselect and then drill down to find a better subset. But we don't want to select too many, particularly ones are not that useful, so anything appearing less than 5% of the time is not really that useful, even if looks very appealing (such as \"populationTotalRanking\", only available for 1.43% of the cities in the devset). The next stage is then write down the list of final relations (file \"ch6_cell10_rels.txt\") and filter the devset to only those triples (file \"ch6_cell10_cities100_dev_filtered.ttl\"). That will also further reduce the computation down the road. Cell #10."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Selected 43 relations\n",
      "Filtered down to 1,278,943 triples\n"
     ]
    }
   ],
   "source": [
    "# CELL 10\n",
    "import re\n",
    "from collections import OrderedDict\n",
    "\n",
    "devset_uris = set(map(lambda line:line.split('\\t')[0], open(\"ch6_cell7_dev_pop.tsv\").readlines()))\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "rel_count = OrderedDict()\n",
    "rel_subj_seen = set()\n",
    "with open(\"ch6_cell4_cities1000_dev_base.ttl\") as base:\n",
    "    for line in base:\n",
    "        subj, verb, obj = triple_re.match(line).groups()\n",
    "        if subj in devset_uris:\n",
    "            rel_subj = \"{}-{}\".format(subj,verb)\n",
    "            if rel_subj not in rel_subj_seen:\n",
    "                rel_count[verb] = rel_count.get(verb, 0) + 1\n",
    "                rel_subj_seen.add(rel_subj)\n",
    "\n",
    "# delete populations\n",
    "del rel_count[\"<http://dbpedia.org/ontology/populationTotal>\"]\n",
    "del rel_count[\"<http://dbpedia.org/ontology/populationMetro>\"]\n",
    "del rel_count[\"<http://dbpedia.org/ontology/populationUrban>\"]\n",
    "del rel_count[\"<http://dbpedia.org/ontology/populationRural>\"]\n",
    "                \n",
    "threshold = len(devset_uris) * 0.05\n",
    "selected_rels = sorted(filter(lambda t:t[1] >= threshold, rel_count.items()), key=lambda t:t[0])\n",
    "\n",
    "with open(\"ch6_cell10_rels.txt\", \"w\") as rels:\n",
    "    rels.write(\"\\n\".join(map(lambda t:t[0], selected_rels))+\"\\n\")\n",
    "\n",
    "selected_rels = set(map(lambda t:t[0], selected_rels))\n",
    "written = 0\n",
    "with open(\"ch6_cell10_cities1000_dev_filtered.ttl\", \"w\") as filtered:\n",
    "    with open(\"ch6_cell4_cities1000_dev_base.ttl\") as base:\n",
    "        for line in base:\n",
    "            subj, verb, obj = triple_re.match(line).groups()\n",
    "            if subj in devset_uris and verb in selected_rels:\n",
    "                filtered.write(line)\n",
    "                written += 1\n",
    "\n",
    "print(\"Selected\", len(selected_rels), \"relations\")\n",
    "print(\"Filtered down to {:,} triples\".format(written))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "From all this process we are now down to 43 relations and 1.3 million triples for 48,000 instances. But we still do not have features, as these relations have multiple instances for them and are unordered. To deal with this situation, we can remember the techniques from Chapter 5 for dealing sets, lists, graphs, etc. For the set-based (unordered) nature of the relations, we can sort them. In this case we can sort them by how frequent is the target value. If the target value is more frequent, it might be more informative to the ML. That leaves then with a list of values for each relation. To represent lists, a common technique is to truncate it to a maximum size. We thus need to see how many times each of these 43 relations is defined for a given entity (Cell #11)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<table><tr><th>Rel</th><th>Counts</th><th>Max</th><th>Avg</th><th>Median</th><th>Min</th></tr><tr><td>area</td><td>4803</td><td>('Saint-Germain-Nuelles', 1)</td><td>1.0</td><td>('Labeuvrière', 1)</td><td>('Austin,_Texas', 1)</td></tr>\n",
       "<tr><td>areaCode</td><td>25417</td><td>('Iserlohn', 8)</td><td>1.14</td><td>('Martin,_Slovakia', 1)</td><td>('Ankara', 1)</td></tr>\n",
       "<tr><td>areaLand</td><td>11669</td><td>('Spotswood,_New_Jersey', 2)</td><td>1.69</td><td>('Reinbeck,_Iowa', 2)</td><td>('Amsterdam', 1)</td></tr>\n",
       "<tr><td>areaTotal</td><td>29917</td><td>('Spotswood,_New_Jersey', 2)</td><td>1.27</td><td>('Osann-Monzel', 1)</td><td>('Ankara', 1)</td></tr>\n",
       "<tr><td>areaWater</td><td>10950</td><td>('Spotswood,_New_Jersey', 2)</td><td>1.42</td><td>('Mebane,_North_Carolina', 1)</td><td>('Amsterdam', 1)</td></tr>\n",
       "<tr><td>birthPlace?inv</td><td>21089</td><td>('London', 6258)</td><td>13.31</td><td>('Itō,_Shizuoka', 3)</td><td>('Kaysersberg', 1)</td></tr>\n",
       "<tr><td>city?inv</td><td>13551</td><td>('London', 709)</td><td>3.9</td><td>('Ashburnham,_Massachusetts', 2)</td><td>('Iquitos', 1)</td></tr>\n",
       "<tr><td>country</td><td>45977</td><td>('Sada,_Galicia', 8)</td><td>1.04</td><td>('Talence', 1)</td><td>('Ankara', 1)</td></tr>\n",
       "<tr><td>deathPlace?inv</td><td>10046</td><td>('London', 3059)</td><td>8.36</td><td>('Longmeadow,_Massachusetts', 2)</td><td>('Sauk_City,_Wisconsin', 1)</td></tr>\n",
       "<tr><td>department</td><td>2587</td><td>('Saint-Denis,_Seine-Saint-Denis', 1)</td><td>1.0</td><td>('Trémuson', 1)</td><td>('Ajaccio', 1)</td></tr>\n",
       "<tr><td>district</td><td>5842</td><td>('Waterloo,_Illinois', 6)</td><td>1.01</td><td>('Müllrose', 1)</td><td>('Aalen', 1)</td></tr>\n",
       "<tr><td>elevation</td><td>28669</td><td>('Córdoba,_Argentina', 4)</td><td>1.33</td><td>('Villaquilambre', 1)</td><td>('Ankara', 1)</td></tr>\n",
       "<tr><td>foundingDate</td><td>3003</td><td>('Tampa,_Florida', 5)</td><td>1.14</td><td>('Rio_Branco,_Acre', 1)</td><td>('Austin,_Texas', 1)</td></tr>\n",
       "<tr><td>ground?inv</td><td>3368</td><td>('Vienna', 134)</td><td>1.94</td><td>('Nödinge-Nol', 1)</td><td>('Kansas_City,_Kansas', 1)</td></tr>\n",
       "<tr><td>headquarter?inv</td><td>2769</td><td>('London', 509)</td><td>5.63</td><td>('Cortona', 1)</td><td>('Wigton', 1)</td></tr>\n",
       "<tr><td>hometown?inv</td><td>4144</td><td>('London', 1562)</td><td>6.52</td><td>('West_Covina,_California', 1)</td><td>('Aston', 1)</td></tr>\n",
       "<tr><td>inseeCode</td><td>3906</td><td>('Saint-Germain-Nuelles', 1)</td><td>1.0</td><td>('Ploumagoar', 1)</td><td>('Ajaccio', 1)</td></tr>\n",
       "<tr><td>isPartOf</td><td>32266</td><td>('Baima,_Baxoi_County', 10)</td><td>2.02</td><td>('Omalur', 2)</td><td>('Amsterdam', 1)</td></tr>\n",
       "<tr><td>isPartOf?inv</td><td>3786</td><td>('Shijiazhuang', 296)</td><td>7.0</td><td>('Santa_Teresa_del_Tuy', 1)</td><td>('Tuxtla_Gutiérrez', 1)</td></tr>\n",
       "<tr><td>leaderName</td><td>3192</td><td>('North_York', 24)</td><td>1.35</td><td>('Pointe-Noire', 1)</td><td>('Ankara', 1)</td></tr>\n",
       "<tr><td>leaderTitle</td><td>12322</td><td>('Canora,_Saskatchewan', 8)</td><td>1.45</td><td>('Bad_Sachsa', 1)</td><td>('Amsterdam', 1)</td></tr>\n",
       "<tr><td>location?inv</td><td>14236</td><td>('London', 2438)</td><td>7.43</td><td>('Schiller_Park,_Illinois', 2)</td><td>('Sèvres', 1)</td></tr>\n",
       "<tr><td>locationCity?inv</td><td>2822</td><td>('London', 684)</td><td>4.33</td><td>('Fall_River,_Massachusetts', 1)</td><td>('Orchard_Park_(town),_New_York', 1)</td></tr>\n",
       "<tr><td>maximumElevation</td><td>4159</td><td>('Zahlé', 2)</td><td>1.0</td><td>('Gières', 1)</td><td>('Aarhus', 1)</td></tr>\n",
       "<tr><td>minimumElevation</td><td>4093</td><td>('Zahlé', 2)</td><td>1.0</td><td>('Raismes', 1)</td><td>('Aarhus', 1)</td></tr>\n",
       "<tr><td>motto</td><td>2689</td><td>('Mérida,_Mérida', 4)</td><td>1.08</td><td>('Moraine,_Ohio', 1)</td><td>('Amsterdam', 1)</td></tr>\n",
       "<tr><td>nearestCity?inv</td><td>4136</td><td>('Middletown,_Orange_County,_New_York', 46)</td><td>1.78</td><td>('Owatonna,_Minnesota', 1)</td><td>('Copenhagen', 1)</td></tr>\n",
       "<tr><td>populationAsOf</td><td>5303</td><td>('Gilroy,_California', 2)</td><td>1.0</td><td>('Ellrich', 1)</td><td>('Aarhus', 1)</td></tr>\n",
       "<tr><td>populationDensity</td><td>13788</td><td>('Batavia,_New_York', 2)</td><td>1.41</td><td>('Santiz', 1)</td><td>('Austin,_Texas', 1)</td></tr>\n",
       "<tr><td>postalCode</td><td>31782</td><td>('Nijemci', 8)</td><td>1.0</td><td>('Lingig,_Surigao_del_Sur', 1)</td><td>('Ankara', 1)</td></tr>\n",
       "<tr><td>region</td><td>4711</td><td>('Centralia,_Illinois', 4)</td><td>1.01</td><td>('La_Châtaigneraie', 1)</td><td>('Aachen', 1)</td></tr>\n",
       "<tr><td>residence?inv</td><td>5283</td><td>('New_York_City', 703)</td><td>4.4</td><td>('Chonburi_(city)', 1)</td><td>('Kiyosu', 1)</td></tr>\n",
       "<tr><td>routeEnd?inv</td><td>4079</td><td>('Chicago', 33)</td><td>1.46</td><td>('Nongstoin', 1)</td><td>('Sleepy_Hollow,_New_York', 1)</td></tr>\n",
       "<tr><td>routeJunction?inv</td><td>5301</td><td>('Chicago', 23)</td><td>1.94</td><td>('Smithville,_Tennessee', 1)</td><td>('Sleepy_Hollow,_New_York', 1)</td></tr>\n",
       "<tr><td>routeStart?inv</td><td>4136</td><td>('Chicago', 16)</td><td>1.53</td><td>('Hilliard,_Florida', 1)</td><td>('Dana_Point,_California', 1)</td></tr>\n",
       "<tr><td>timeZone</td><td>28615</td><td>('Jerusalem', 4)</td><td>1.35</td><td>('Bhikhi,_Punjab', 1)</td><td>('Ames,_Iowa', 1)</td></tr>\n",
       "<tr><td>type</td><td>21497</td><td>('Kandi,_Benin', 3)</td><td>1.02</td><td>('Jericho,_Vermont', 1)</td><td>('Ankara', 1)</td></tr>\n",
       "<tr><td>utcOffset</td><td>27762</td><td>('Lőrinci', 3)</td><td>1.74</td><td>('Marietta,_Ohio', 2)</td><td>('Algiers', 1)</td></tr>\n",
       "<tr><td>22-rdf-syntax-ns#type</td><td>48868</td><td>('Okny', 1)</td><td>1.0</td><td>('Labarthe-Rivière', 1)</td><td>('Ankara', 1)</td></tr>\n",
       "<tr><td>rdf-schema#seeAlso</td><td>3426</td><td>('Dubai', 19)</td><td>2.69</td><td>('Coventry,_Rhode_Island', 3)</td><td>('Bishkek', 1)</td></tr>\n",
       "<tr><td>homepage</td><td>20319</td><td>('Pune', 4)</td><td>1.0</td><td>('Nueva_Loja', 1)</td><td>('Amsterdam', 1)</td></tr>\n",
       "<tr><td>name</td><td>45193</td><td>('Songtao_Miao_Autonomous_County', 5)</td><td>1.19</td><td>('Castillejo_de_Martín_Viejo', 1)</td><td>('Ankara', 1)</td></tr>\n",
       "<tr><td>nick</td><td>2764</td><td>('Davao_City', 10)</td><td>1.1</td><td>('Rotterdam_(town),_New_York', 1)</td><td>('Ankara', 1)</td></tr></table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 11\n",
    "import re\n",
    "from collections import OrderedDict\n",
    "\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "rel_count_per_city = OrderedDict()\n",
    "with open(\"ch6_cell10_cities1000_dev_filtered.ttl\") as base:\n",
    "    for line in base:\n",
    "        subj, verb, obj = triple_re.match(line).groups()\n",
    "        if verb not in rel_count_per_city:\n",
    "            rel_count_per_city[verb] = OrderedDict()\n",
    "        rel_count_per_city[verb][subj] = rel_count_per_city[verb].get(subj, 0) + 1\n",
    "\n",
    "# sort and compute min, max, average and median\n",
    "table = list()\n",
    "for rel in sorted(rel_count_per_city.keys()):\n",
    "    city_counts = sorted(rel_count_per_city[rel].items(), key=lambda t:t[1])\n",
    "    city_counts = list(map(lambda t:(t[0][:-1].split(\"/\")[-1],t[1]), city_counts))\n",
    "    table.append( (rel[:-1].split(\"/\")[-1], # rel \n",
    "                   len(city_counts), # counts\n",
    "          city_counts[-1], # max\n",
    "          int(sum(map(lambda t:t[1], city_counts)) / len(city_counts) * 100) / 100, # avg\n",
    "          city_counts[int(len(city_counts) / 2)], # median\n",
    "          city_counts[0])  # min  \n",
    "    )\n",
    "from IPython.display import HTML, display\n",
    "display(HTML(\"<table><tr><th>Rel</th><th>Counts</th><th>Max</th><th>Avg</th><th>Median</th><th>Min</th></tr>\" +\n",
    "            \"\\n\".join(list(map(lambda r:\"<tr><td>\"+\"</td><td>\".join(list(map(str,r)))+\n",
    "                               \"</td></tr>\", table))) +\"</table>\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "There are multiple things we can see here. First, no relation has a minimum that it is not 1. That is good because it means no relation is inherently multiple. We can also see that the inverted relations have substantive maximums, which make sense that cities such as London will be the birthplaces of thousands of Wikipedia-worthy people. Clearly the most important signal here is the number of such inbound relations rather than the actual entities being referred, but we will leave that to be brought about by some feature selection techniques as the ones discussed in Chapter 4. Some the averages are also pretty high but they seem to be driven by few megacities. The medians are much informative, with only a few beyond one. I will thus focus on the medians and extract the file \"ch6_cell12_rels_median_fan_outs.tsv\" (Cell #12)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "# CELL 12\n",
    "import re\n",
    "from collections import OrderedDict\n",
    "\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "rel_count_per_city = OrderedDict()\n",
    "with open(\"ch6_cell10_cities1000_dev_filtered.ttl\") as base:\n",
    "    for line in base:\n",
    "        subj, verb, obj = triple_re.match(line).groups()\n",
    "        if verb not in rel_count_per_city:\n",
    "            rel_count_per_city[verb] = OrderedDict()\n",
    "        rel_count_per_city[verb][subj] = rel_count_per_city[verb].get(subj, 0) + 1\n",
    "\n",
    "# sort and compute min, max, average and median\n",
    "with open(\"ch6_cell12_rels_median_fan_outs.tsv\", \"w\") as medians:\n",
    "    for rel in sorted(rel_count_per_city.keys()):\n",
    "        city_counts = sorted(rel_count_per_city[rel].items(), key=lambda t:t[1])\n",
    "        medians.write(\"{}\\t{}\\n\".format(rel,city_counts[int(len(city_counts) / 2)][1]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Base Featurization\n",
    "\n",
    "We are now almost ready to produce a base featurization, all we are missing is to have global counts for all literals and entities that are targets of the relations and use them to sort and select the top ones according to the median fan-out computed in Cell #11. To such features I will also add the total count per relation and a large total of all relevant relations. Missing features are marked by the string \"N/A\". Adding the population produces the file \"ch6_cell13_dev_feat0.tsv\", the first full featurization of the dataset (Cell #13)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Rows written (including header): 48,869\n"
     ]
    }
   ],
   "source": [
    "# CELL 13\n",
    "import re\n",
    "from collections import OrderedDict\n",
    "\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "# compute counts for ordering values\n",
    "popularity = dict()\n",
    "with open(\"ch6_cell10_cities1000_dev_filtered.ttl\") as base:\n",
    "    for line in base:\n",
    "        subj, verb, obj = triple_re.match(line).groups()\n",
    "        popularity[obj] = popularity.get(obj, 0) + 1\n",
    "\n",
    "# fan outs\n",
    "fan_outs = dict()\n",
    "with open(\"ch6_cell12_rels_median_fan_outs.tsv\") as medians:\n",
    "    for line in medians:\n",
    "        rel, fan_out = line.strip().split(\"\\t\")\n",
    "        fan_outs[rel] = int(fan_out)\n",
    "rels = sorted(fan_outs.keys())\n",
    "        \n",
    "# population\n",
    "pops = dict()\n",
    "with open(\"ch6_cell7_dev_pop.tsv\") as dev_pop:\n",
    "    for line in dev_pop:\n",
    "        city, pop = line.strip().split(\"\\t\")\n",
    "        pops[city] = float(pop)\n",
    "        \n",
    "# full graph\n",
    "graph = OrderedDict()\n",
    "with open(\"ch6_cell10_cities1000_dev_filtered.ttl\") as base:\n",
    "    for line in base:\n",
    "        subj, verb, obj = triple_re.match(line).groups()\n",
    "        if verb in fan_outs:\n",
    "            if subj not in graph:\n",
    "                graph[subj] = OrderedDict()\n",
    "            if verb not in graph[subj]:\n",
    "                graph[subj][verb] = list()\n",
    "            graph[subj][verb].append(obj)\n",
    "\n",
    "# putting it all together\n",
    "rows_written = 0\n",
    "with open(\"ch6_cell13_dev_feat0.tsv\", \"w\") as feats:\n",
    "    # header\n",
    "    feats.write(\"name\\trel#count\")\n",
    "    rel_start = dict()\n",
    "    idx = 2\n",
    "    for rel in rels:\n",
    "        fan_out = fan_outs[rel]\n",
    "        rel_start[rel] = idx\n",
    "        rel = rel[1:-1]\n",
    "        feats.write(\"\\t{}#count\".format(rel))\n",
    "        for i in range(fan_out):\n",
    "            feats.write(\"\\t{}#{}\".format(rel, i+1))\n",
    "        idx += fan_out + 1\n",
    "    feats.write(\"\\tpopulation\\n\")\n",
    "    rows_written += 1\n",
    "    \n",
    "    # data\n",
    "    for city, data in graph.items():\n",
    "        feats.write(\"{}\\t{}\".format(city, sum(map(len,data.values()))))\n",
    "        for rel in rels:\n",
    "            fan_out = fan_outs[rel]\n",
    "            if rel not in data:\n",
    "                feats.write(\"\\tNA\" * (fan_out + 1))\n",
    "            else:\n",
    "                feats.write(\"\\t{}\".format(len(data[rel])))\n",
    "                values = sorted(data[rel], key=lambda o:popularity[o], reverse=True)[:fan_out]\n",
    "                feats.write(\"\\t\" + \"\\t\".join(values))\n",
    "                written = len(values)\n",
    "                while written < fan_out:\n",
    "                    feats.write(\"\\tNA\")\n",
    "                    written += 1\n",
    "        feats.write(\"\\t{}\\n\".format(pops[city]))\n",
    "        rows_written += 1\n",
    "print(\"Rows written (including header): {:,}\".format(rows_written))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We are now at the moment to start thinking about the ML algorithm we will employ over this data. It contains almost 100 columns of different types and a numeric target. Let us start by plotting the target and its distribution (Cell #14)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<module 'matplotlib.pyplot' from '/vol/bulk/work/feateng/finaltest/chapter6/feateng/lib/python3.7/site-packages/matplotlib/pyplot.py'>"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 14\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "\n",
    "pops = list()\n",
    "with open(\"ch6_cell13_dev_feat0.tsv\") as feats:\n",
    "    next(feats) # skip header\n",
    "    for line in feats:\n",
    "        pops.append(float(line.strip().split(\"\\t\")[-1]) / 10e6)\n",
    "pops = sorted(pops)\n",
    "\n",
    "plt.plot(pops, label=\"population\", color='gray')\n",
    "plt.ylabel('population in millions')\n",
    "plt.savefig(\"ch6_cell14_population.pdf\", bbox_inches='tight', dpi=300)\n",
    "plt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "That curve is quite steep at the end. To flatten it down, we can apply a logarithmic function as it helps put such growths processes in perspective, as discussed in Chapter 3. Instead of the default logarithm on base $e$ I will use the more intuitive base 10 that tell us the number of digits in the population for a city. Cell #15 shows the result."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<module 'matplotlib.pyplot' from '/vol/bulk/work/feateng/finaltest/chapter6/feateng/lib/python3.7/site-packages/matplotlib/pyplot.py'>"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 15\n",
    "import math\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "\n",
    "pops = list()\n",
    "with open(\"ch6_cell13_dev_feat0.tsv\") as feats:\n",
    "    next(feats) # skip header\n",
    "    for line in feats:\n",
    "        pop = float(line.strip().split(\"\\t\")[-1])\n",
    "        if pop == 0:\n",
    "            pop = 1\n",
    "        pops.append(math.log(pop, 10))\n",
    "pops = sorted(pops)\n",
    "\n",
    "plt.plot(pops, label=\"log population\", color='gray')\n",
    "plt.ylabel('log population in millions')\n",
    "plt.savefig(\"ch6_cell15_log_population.pdf\", bbox_inches='tight', dpi=300)\n",
    "plt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The first thing to notice in this figure is that there seems to be many cities in the data set with less that 1,000 inhabitants. Some even with only 1 person! Looking a little bit into the triples, I can see that some are actual mistakes in Wikipedia (a city with 1,000,000 people got typed as 1.000.000, etc). Other are small towns and hamlets that are less than a 1,000 people (but might have been more than a 1,000 at the time GeoNames built their list). Following the discussion on outliers in Chapter 2, these are outliers that are worth removing from the data. I will still leave them in the held-out final test data as their impact in production needs to be seen and evaluated.\n",
    "\n",
    "We thus remove these outliers and produce a filtered file \"ch6_cell16_dev_feat0_filtered.tsv\" (Cell #16)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Rows written (including header): 44,959\n"
     ]
    }
   ],
   "source": [
    "# CELL 16\n",
    "written = 0\n",
    "with open(\"ch6_cell16_dev_feat0_filtered.tsv\", \"w\") as filtered:\n",
    "    with open(\"ch6_cell13_dev_feat0.tsv\") as feats:\n",
    "        filtered.write(next(feats))\n",
    "        for line in feats:\n",
    "            pop = float(line.strip().split(\"\\t\")[-1])\n",
    "            if pop >= 1000:\n",
    "                filtered.write(line)\n",
    "                written += 1\n",
    "print(\"Rows written (including header): {:,}\".format(written))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "So we did not throw a lot of data with this change, good.\n",
    "\n",
    "Now we can try some learning for EDA. Let us try some regression models on it against the single feature of \"number of relations defined\". Let's start with SVR using RBFs (Cell #17)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Training on 8,963 cities\n",
      "RMSE 0.617989060494947\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x7f92c6cdb5d0>"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x360 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 17\n",
    "import random\n",
    "import math\n",
    "from sklearn.svm import SVR\n",
    "import numpy as np\n",
    "\n",
    "rand = random.Random(42)\n",
    "train_data = list()\n",
    "test_data = list()\n",
    "with open(\"ch6_cell16_dev_feat0_filtered.tsv\") as feats:\n",
    "    next(feats) # skip header\n",
    "    for line in feats:\n",
    "        fields = line.strip().split(\"\\t\")\n",
    "        pop = float(fields[-1])\n",
    "        rels = float(fields[1])\n",
    "        row = (math.log(rels, 10), math.log(pop, 10))\n",
    "        if rand.random() < 0.8:\n",
    "            test_data.append(row) \n",
    "        else:\n",
    "            train_data.append(row)\n",
    "               \n",
    "test_data = sorted(test_data, key=lambda t:t[1])\n",
    "\n",
    "train_data = np.array(train_data)\n",
    "test_data = np.array(test_data)\n",
    "\n",
    "xtrain = train_data[:,0]\n",
    "ytrain = train_data[:,1]\n",
    "xtest = test_data[:,0]\n",
    "ytest = test_data[:,1]\n",
    "\n",
    "# SVRs need scaling\n",
    "xtrain_min = xtrain.min(); xtrain_max = xtrain.max()\n",
    "xtrain_scaling = 1.0 / (xtrain_max - xtrain_min)\n",
    "xtrain -= xtrain_min\n",
    "xtrain *= xtrain_scaling\n",
    "\n",
    "ytrain_min = ytrain.min(); ytrain_max = ytrain.max()\n",
    "ytrain_scaling = 1.0 / (ytrain_max - ytrain_min)\n",
    "ytrain -= ytrain_min\n",
    "ytrain *= ytrain_scaling\n",
    "\n",
    "xtest -= xtrain_min\n",
    "xtest *= xtrain_scaling\n",
    "ytest_orig = ytest.copy()\n",
    "ytest -= ytrain_min\n",
    "ytest *= ytrain_scaling\n",
    "\n",
    "# train\n",
    "print(\"Training on {:,} cities\".format(len(xtrain)))\n",
    "best_c = 100.0\n",
    "best_epsilon = 0.1\n",
    "PARAM_DO_GRID_SEARCH = False\n",
    "      \n",
    "if PARAM_DO_GRID_SEARCH:\n",
    "    best_rmse = 1000\n",
    "    for c in [0.01, 0.1, 1.0, 2.0, 5.0, 10.0, 50.0, 100.0]:\n",
    "        svr_rbf = SVR(epsilon=0.05, C=c, gamma='auto')\n",
    "        svr_rbf.fit(xtrain.reshape(-1,1), ytrain)\n",
    "        ytest_pred = svr_rbf.predict(xtest.reshape(-1,1))\n",
    "        ytest_pred *= 1.0/ytrain_scaling\n",
    "        ytest_pred += ytrain_min\n",
    "        RMSE = math.sqrt(sum((ytest_orig - ytest_pred)**2) / len(ytest))\n",
    "        print(\"C\", c, \"RMSE\", RMSE)\n",
    "        if RMSE < best_rmse:\n",
    "            best_c = c\n",
    "            best_rmse = RMSE\n",
    "\n",
    "    print(\"Best C\", best_c,\"best RMSE\", best_rmse)        \n",
    "\n",
    "    best_rmse = 1000\n",
    "    for epsilon in [0.0001, 0.001, 0.01, 0.05, 0.1, 0.5, 0.0, 1.0, 10.0, 100.0]:\n",
    "        svr_rbf = SVR(epsilon=epsilon, C=best_c, gamma='auto')\n",
    "        svr_rbf.fit(xtrain.reshape(-1,1), ytrain)\n",
    "        ytest_pred = svr_rbf.predict(xtest.reshape(-1,1))\n",
    "        ytest_pred *= 1.0/ytrain_scaling\n",
    "        ytest_pred += ytrain_min\n",
    "        RMSE = math.sqrt(sum((ytest_orig - ytest_pred)**2) / len(ytest))\n",
    "        print(\"Epsilon\", epsilon, \"RMSE\", RMSE)\n",
    "        if RMSE < best_rmse:\n",
    "            best_epsilon = epsilon\n",
    "            best_rmse = RMSE\n",
    "\n",
    "    print(\"Best epsilon\", best_epsilon,\"best RMSE\", best_rmse)\n",
    "svr_rbf = SVR(epsilon=best_epsilon, C=best_c, gamma='auto')\n",
    "svr_rbf.fit(xtrain.reshape(-1,1), ytrain)\n",
    "ytest_pred = svr_rbf.predict(xtest.reshape(-1,1))\n",
    "ytest_pred *= 1.0/ytrain_scaling\n",
    "ytest_pred += ytrain_min\n",
    "RMSE = math.sqrt(sum((ytest_orig - ytest_pred)**2) / len(ytest))\n",
    "print(\"RMSE\", RMSE)\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "plt.rcParams['figure.figsize'] = [20, 5]\n",
    "plt.plot(ytest_pred, label=\"predicted\", color='gray')\n",
    "plt.plot(ytest_orig, label=\"actual\", color='black')\n",
    "plt.ylabel('scaled log population')\n",
    "plt.savefig(\"ch6_cell17_svr_one_feat.pdf\", bbox_inches='tight', dpi=300)\n",
    "plt.legend()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "While the RMSE is not that bad for only one feature (less that one order of magnitude) and the flat bottom of the figure can be explained by the fact the SVR has only a single support vector (from its single feature), the figure still does not seem to track the curve that well. To further see if this is an issue with SVR or with the feature, I plotted a 20% sample of the total number of relations against the log of the population in Cell #18."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<module 'matplotlib.pyplot' from '/vol/bulk/work/feateng/finaltest/chapter6/feateng/lib/python3.7/site-packages/matplotlib/pyplot.py'>"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 720x720 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 18\n",
    "import math\n",
    "import random\n",
    "\n",
    "pops = list()\n",
    "num_rels = list()\n",
    "rand = random.Random(42)\n",
    "with open(\"ch6_cell16_dev_feat0_filtered.tsv\") as feats:\n",
    "    next(feats) # skip header\n",
    "    for line in feats:\n",
    "        fields = line.strip().split(\"\\t\")\n",
    "        pop = float(fields[-1])\n",
    "        rels = float(fields[1])\n",
    "        if rand.random() < 0.2:\n",
    "            pops.append(math.log(pop, 10))\n",
    "            num_rels.append(math.log(rels, 10))\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "plt.rcParams['figure.figsize'] = [10, 10]\n",
    "plt.plot(num_rels, pops, '.', color='gray')\n",
    "plt.ylabel('scaled log population')\n",
    "plt.xlabel('scaled number of relations')\n",
    "plt.savefig(\"ch6_cell18_logpop_vs_log_items.pdf\", bbox_inches='tight', dpi=300)\n",
    "plt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "It seems the correlation is there but it is weak and SVR is extracting as good information as it is. I thus settle for SVR and we are ready for our first feature set.\n",
    "\n",
    "## First Featurization\n",
    "\n",
    "As SVR only takes numeric features, the first step is to typify the raw columns we have and see which ones are quasi-numeric (that is, numeric 80% of the time) and whether the ones that are not have a small (100 or less) unique values. This typification is in Cell #19."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Total cols 99\n",
      "Numeric cols 54\n",
      "Categorical cols\n",
      "   http://dbpedia.org/ontology/utcOffset#1\n",
      "   http://dbpedia.org/ontology/utcOffset#2\n",
      "   http://www.w3.org/1999/02/22-rdf-syntax-ns#type#1\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<table><tr><th>Rel</th><th>Counts</th></tr><tr><td>name</td><td>44959</td></tr>\n",
       "<tr><td>areaCode#1</td><td>5667</td></tr>\n",
       "<tr><td>birthPlace?inv#1</td><td>20183</td></tr>\n",
       "<tr><td>birthPlace?inv#2</td><td>13748</td></tr>\n",
       "<tr><td>birthPlace?inv#3</td><td>10523</td></tr>\n",
       "<tr><td>city?inv#1</td><td>11381</td></tr>\n",
       "<tr><td>city?inv#2</td><td>6884</td></tr>\n",
       "<tr><td>country#1</td><td>253</td></tr>\n",
       "<tr><td>deathPlace?inv#1</td><td>9848</td></tr>\n",
       "<tr><td>deathPlace?inv#2</td><td>5215</td></tr>\n",
       "<tr><td>department#1</td><td>172</td></tr>\n",
       "<tr><td>district#1</td><td>1182</td></tr>\n",
       "<tr><td>foundingDate#1</td><td>2649</td></tr>\n",
       "<tr><td>ground?inv#1</td><td>3284</td></tr>\n",
       "<tr><td>headquarter?inv#1</td><td>2671</td></tr>\n",
       "<tr><td>hometown?inv#1</td><td>4030</td></tr>\n",
       "<tr><td>inseeCode#1</td><td>3855</td></tr>\n",
       "<tr><td>isPartOf#1</td><td>2761</td></tr>\n",
       "<tr><td>isPartOf#2</td><td>7640</td></tr>\n",
       "<tr><td>isPartOf?inv#1</td><td>3331</td></tr>\n",
       "<tr><td>leaderName#1</td><td>1839</td></tr>\n",
       "<tr><td>leaderTitle#1</td><td>454</td></tr>\n",
       "<tr><td>location?inv#1</td><td>13017</td></tr>\n",
       "<tr><td>location?inv#2</td><td>7984</td></tr>\n",
       "<tr><td>locationCity?inv#1</td><td>2704</td></tr>\n",
       "<tr><td>motto#1</td><td>2587</td></tr>\n",
       "<tr><td>nearestCity?inv#1</td><td>3534</td></tr>\n",
       "<tr><td>populationAsOf#1</td><td>259</td></tr>\n",
       "<tr><td>postalCode#1</td><td>25603</td></tr>\n",
       "<tr><td>region#1</td><td>404</td></tr>\n",
       "<tr><td>residence?inv#1</td><td>5033</td></tr>\n",
       "<tr><td>routeEnd?inv#1</td><td>3968</td></tr>\n",
       "<tr><td>routeJunction?inv#1</td><td>2714</td></tr>\n",
       "<tr><td>routeStart?inv#1</td><td>4019</td></tr>\n",
       "<tr><td>timeZone#1</td><td>178</td></tr>\n",
       "<tr><td>type#1</td><td>582</td></tr>\n",
       "<tr><td>rdf-schema#seeAlso#1</td><td>896</td></tr>\n",
       "<tr><td>rdf-schema#seeAlso#2</td><td>692</td></tr>\n",
       "<tr><td>rdf-schema#seeAlso#3</td><td>423</td></tr>\n",
       "<tr><td>homepage#1</td><td>18912</td></tr>\n",
       "<tr><td>name#1</td><td>40080</td></tr>\n",
       "<tr><td>nick#1</td><td>2601</td></tr></table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 19\n",
    "types = dict()\n",
    "data = list()\n",
    "header = None\n",
    "with open(\"ch6_cell16_dev_feat0_filtered.tsv\") as feats:\n",
    "    header = next(feats).strip().split('\\t')\n",
    "    for line in feats:\n",
    "        data.append( line.strip().split('\\t'))\n",
    "\n",
    "# numeric columns\n",
    "num_cols = set()\n",
    "for idx in range(len(header)):\n",
    "    total_non_na = 0\n",
    "    is_num = 0\n",
    "    for row in data:\n",
    "        if row[idx] == 'NA':\n",
    "            continue\n",
    "        total_non_na += 1\n",
    "        row_val = row[idx]\n",
    "        if '\"^^' in row_val:\n",
    "            row_val = row_val[1:row_val.index('\"^^')]\n",
    "        try:\n",
    "            float(row_val)\n",
    "            is_num += 1\n",
    "        except ValueError:\n",
    "            pass\n",
    "    if is_num > total_non_na * 0.8:\n",
    "        num_cols.add(idx)\n",
    "        \n",
    "print(\"Total cols\", len(header))\n",
    "print(\"Numeric cols\", len(num_cols))\n",
    "\n",
    "# differentiate categorical from free text, up to 100 categories\n",
    "cat_cols = dict()\n",
    "others = list()\n",
    "for idx in range(len(header)):\n",
    "    if idx in num_cols:\n",
    "        continue\n",
    "    values = set()\n",
    "    for row in data:\n",
    "        if row[idx] == 'NA':\n",
    "            continue\n",
    "        values.add(row[idx])\n",
    "    if len(values) < 100:\n",
    "        cat_cols[idx] = values\n",
    "    else:\n",
    "        others.append( (header[idx], len(values)) )\n",
    "\n",
    "print(\"Categorical cols\")\n",
    "for col in sorted(cat_cols.keys()):\n",
    "        print(\"   \" + header[col])\n",
    "from IPython.display import HTML, display\n",
    "display(HTML(\"<table><tr><th>Rel</th><th>Counts</th></tr>\" +\n",
    "            \"\\n\".join(list(map(lambda r: \"<tr><td>{}</td><td>{}</td></tr>\".format(r[0].split(\"/\")[-1], r[1]), \n",
    "                               others))) +\"</table>\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Therefore, of the 99 columns, 54 are numeric but only 3 can be captured by 100 unique values. As we can expect more of them to be quasi-categorical, I employ a technique for category discretization discussed in Chapter 2, introducing an OTHER category to handle the rest of the values if they account for less than 20% of the total number of values (Cell #20). \n",
    "\n",
    "I also drop constant columns (which are uninformative) netting 50 numeric columns and 11 categorical. The remaining columns have a variety that cannot be discretized. These columns will require text processing or they are just uninformative for this approach. The selected categorical relations with their total number of categories is shown. Types go to \"ch6_cell20_feat0_types.tsv\"."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Total cols 99\n",
      "Constant cols 4\n",
      "Numeric cols 50\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<h3>Categorical cols (11)</h3><table><tr><th>Rel</th><th>Categories</th></tr><tr><td>country#1</td><td>21</td></tr>\n",
       "<tr><td>department#1</td><td>59</td></tr>\n",
       "<tr><td>leaderTitle#1</td><td>4</td></tr>\n",
       "<tr><td>populationAsOf#1</td><td>15</td></tr>\n",
       "<tr><td>region#1</td><td>100</td></tr>\n",
       "<tr><td>timeZone#1</td><td>14</td></tr>\n",
       "<tr><td>type#1</td><td>29</td></tr>\n",
       "<tr><td>utcOffset#1</td><td>7</td></tr>\n",
       "<tr><td>utcOffset#2</td><td>6</td></tr>\n",
       "<tr><td>22-rdf-syntax-ns#type#1</td><td>3</td></tr>\n",
       "<tr><td>rdf-schema#seeAlso#3</td><td>73</td></tr></table><h3>Remaining cols (34)</h3><table><tr><th>Rel</th><th>Counts</th><th>% Covered @ 100</th></tr><tr><td>name</td><td>44959</td><td>0.22</td></tr>\n",
       "<tr><td>areaCode#1</td><td>5667</td><td>26.65</td></tr>\n",
       "<tr><td>birthPlace?inv#1</td><td>20183</td><td>2.07</td></tr>\n",
       "<tr><td>birthPlace?inv#2</td><td>13748</td><td>1.0</td></tr>\n",
       "<tr><td>birthPlace?inv#3</td><td>10523</td><td>1.0</td></tr>\n",
       "<tr><td>city?inv#1</td><td>11381</td><td>5.1</td></tr>\n",
       "<tr><td>city?inv#2</td><td>6884</td><td>3.32</td></tr>\n",
       "<tr><td>deathPlace?inv#1</td><td>9848</td><td>1.52</td></tr>\n",
       "<tr><td>deathPlace?inv#2</td><td>5215</td><td>1.99</td></tr>\n",
       "<tr><td>district#1</td><td>1182</td><td>31.45</td></tr>\n",
       "<tr><td>foundingDate#1</td><td>2649</td><td>10.33</td></tr>\n",
       "<tr><td>ground?inv#1</td><td>3284</td><td>4.09</td></tr>\n",
       "<tr><td>headquarter?inv#1</td><td>2671</td><td>6.47</td></tr>\n",
       "<tr><td>hometown?inv#1</td><td>4030</td><td>4.35</td></tr>\n",
       "<tr><td>inseeCode#1</td><td>3855</td><td>2.79</td></tr>\n",
       "<tr><td>isPartOf#1</td><td>2761</td><td>56.1</td></tr>\n",
       "<tr><td>isPartOf#2</td><td>7640</td><td>15.33</td></tr>\n",
       "<tr><td>isPartOf?inv#1</td><td>3331</td><td>11.27</td></tr>\n",
       "<tr><td>leaderName#1</td><td>1839</td><td>39.47</td></tr>\n",
       "<tr><td>location?inv#1</td><td>13017</td><td>3.38</td></tr>\n",
       "<tr><td>location?inv#2</td><td>7984</td><td>3.06</td></tr>\n",
       "<tr><td>locationCity?inv#1</td><td>2704</td><td>7.36</td></tr>\n",
       "<tr><td>motto#1</td><td>2587</td><td>6.18</td></tr>\n",
       "<tr><td>nearestCity?inv#1</td><td>3534</td><td>11.07</td></tr>\n",
       "<tr><td>postalCode#1</td><td>25603</td><td>3.27</td></tr>\n",
       "<tr><td>residence?inv#1</td><td>5033</td><td>4.76</td></tr>\n",
       "<tr><td>routeEnd?inv#1</td><td>3968</td><td>3.49</td></tr>\n",
       "<tr><td>routeJunction?inv#1</td><td>2714</td><td>14.05</td></tr>\n",
       "<tr><td>routeStart?inv#1</td><td>4019</td><td>3.09</td></tr>\n",
       "<tr><td>rdf-schema#seeAlso#1</td><td>896</td><td>74.52</td></tr>\n",
       "<tr><td>rdf-schema#seeAlso#2</td><td>692</td><td>76.44</td></tr>\n",
       "<tr><td>homepage#1</td><td>18912</td><td>2.89</td></tr>\n",
       "<tr><td>name#1</td><td>40080</td><td>1.1</td></tr>\n",
       "<tr><td>nick#1</td><td>2601</td><td>7.98</td></tr></table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 20\n",
    "from collections import OrderedDict\n",
    "\n",
    "types = dict()\n",
    "data = list()\n",
    "header = None\n",
    "with open(\"ch6_cell16_dev_feat0_filtered.tsv\") as feats:\n",
    "    header = next(feats).strip().split('\\t')\n",
    "    for line in feats:\n",
    "        data.append( line.strip().split('\\t'))\n",
    "\n",
    "# constant columns\n",
    "const_cols = set()\n",
    "\n",
    "for idx in range(len(header)):\n",
    "    value = None\n",
    "    constant = True\n",
    "    for row in data:\n",
    "        if row[idx] == 'NA':\n",
    "            continue\n",
    "        if value is None:\n",
    "            value = row[idx]\n",
    "        elif value != row[idx]:\n",
    "            constant = False\n",
    "            break\n",
    "    if constant:\n",
    "        const_cols.add(idx)\n",
    "\n",
    "# integer columns\n",
    "num_cols = set()\n",
    "for idx in range(len(header)):\n",
    "    if idx in const_cols:\n",
    "        continue\n",
    "    total_non_na = 0\n",
    "    is_num = 0\n",
    "    for row in data:\n",
    "        if row[idx] == 'NA':\n",
    "            continue\n",
    "        total_non_na += 1\n",
    "        row_val = row[idx]\n",
    "        if '\"^^' in row_val:\n",
    "            row_val = row_val[1:row_val.index('\"^^')]\n",
    "        try:\n",
    "            float(row_val)\n",
    "            is_num += 1\n",
    "        except ValueError:\n",
    "            pass\n",
    "    if is_num > total_non_na * 0.8:\n",
    "        num_cols.add(idx)\n",
    "        \n",
    "print(\"Total cols\", len(header))\n",
    "print(\"Constant cols\", len(const_cols))\n",
    "print(\"Numeric cols\", len(num_cols))\n",
    "\n",
    "# differentiate categorical from free text, up to 100 categories, including \"other\"\n",
    "cat_cols = dict()\n",
    "others = list()\n",
    "for idx in range(len(header)):\n",
    "    if idx in num_cols or idx in const_cols:\n",
    "        continue\n",
    "    values = OrderedDict()\n",
    "    total_obs = 0\n",
    "    for row in data:\n",
    "        if row[idx] == 'NA':\n",
    "            continue\n",
    "        values[row[idx]] = values.get(row[idx], 0) + 1\n",
    "        total_obs += 1\n",
    "    values = sorted(values.items(), key=lambda x:x[1])\n",
    "    total_values = len(values)\n",
    "    selected_values = list()\n",
    "    selected_obs = 0\n",
    "    while selected_obs * 1.0 / total_obs < 0.8 and len(selected_values) < 100:\n",
    "        value, count = values.pop()\n",
    "        selected_values.append(value)\n",
    "        selected_obs += count\n",
    "    if len(selected_values) < 100:\n",
    "        selected_values.append('OTHER')\n",
    "        cat_cols[idx] = selected_values\n",
    "    else:\n",
    "        others.append( (header[idx], total_values, int(selected_obs * 1.0 / total_obs * 10000)/100) )\n",
    "        \n",
    "with open(\"ch6_cell20_feat0_types.tsv\", \"w\") as types:\n",
    "    for idx in range(len(header)):\n",
    "        types.write(\"{}\\t{}\".format(idx, header[idx]))\n",
    "        if idx in num_cols:\n",
    "            types.write(\"\\tNUMERIC\\n\")\n",
    "        elif idx in cat_cols:\n",
    "            types.write(\"\\t\" + \"\\t\".join(cat_cols[idx]) + \"\\n\")\n",
    "        else:\n",
    "            types.write(\"\\tIGNORE\\n\")\n",
    "        \n",
    "table1 = (\"<table><tr><th>Rel</th><th>Categories</th></tr>\" +\n",
    "   \"\\n\".join(list(map(lambda x: \"<tr><td>{}</td><td>{}</td></tr>\".format(header[x].split(\"/\")[-1], \n",
    "                                                                         len(cat_cols[x])), cat_cols.keys()))) +\n",
    "    \"</table>\")\n",
    "table2 = (\"<table><tr><th>Rel</th><th>Counts</th><th>% Covered @ 100</th></tr>\" +\n",
    "            \"\\n\".join(list(map(lambda r: \n",
    "                               \"<tr><td>{}</td><td>{}</td><td>{}</td></tr>\".format(r[0].split(\"/\")[-1], \n",
    "                                                                                   r[1], r[2]), \n",
    "                               others))) +\"</table>\")\n",
    "from IPython.display import HTML, display\n",
    "display(HTML(\"<h3>Categorical cols ({})</h3>\".format(len(cat_cols)) + \n",
    "             table1 + \"<h3>Remaining cols ({})</h3>\".format(len(others)) + table2))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "I now proceed to perform a one hot encoding of the categories (cf. Chapter 3) and imputing the values to the OTHER for categorical (cf. Chapter 2) or to 0.0 for numeric into file \"ch6_cell21_dev_feat1.tsv\" (Cell #21). Its vector size has 380 features plus the target variable."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Vector size 381\n"
     ]
    }
   ],
   "source": [
    "# CELL 21\n",
    "from collections import OrderedDict\n",
    "\n",
    "types = list(map(lambda line: line.strip().split(\"\\t\"), \n",
    "                 open(\"ch6_cell20_feat0_types.tsv\").readlines()))\n",
    "header = None\n",
    "data = list()\n",
    "with open(\"ch6_cell16_dev_feat0_filtered.tsv\") as feats:\n",
    "    header = next(feats).strip().split('\\t')\n",
    "    for line in feats:\n",
    "        data.append( line.strip().split('\\t'))\n",
    "\n",
    "total_size = 0\n",
    "cat_to_pos = dict()\n",
    "for type_line in types:\n",
    "    if type_line[2] == 'NUMERIC':\n",
    "        total_size += 1\n",
    "    elif type_line[2] != 'IGNORE':\n",
    "        cat_to_pos[int(type_line[0])] = dict(zip(type_line[2:], range(len(type_line) - 2)))\n",
    "        total_size += len(type_line) - 2\n",
    "    \n",
    "with open(\"ch6_cell21_dev_feat1.tsv\", \"w\") as feat1:\n",
    "    feat1.write('name') # add name for error analysis\n",
    "    for type_line in types:\n",
    "        if type_line[2] == 'NUMERIC':\n",
    "            feat1.write('\\t' + type_line[1])\n",
    "        elif type_line[2] != 'IGNORE':\n",
    "            for value in type_line[2:]:\n",
    "                feat1.write('\\t{}@{}'.format(type_line[1], value))\n",
    "    feat1.write('\\n')\n",
    "    for row in data:\n",
    "        v = ['NA',] * (total_size + 1)\n",
    "        v[0] = row[0] # add name for error analysis\n",
    "        current = 1\n",
    "        for type_line in types:\n",
    "            idx = int(type_line[0])\n",
    "            value = row[idx]\n",
    "            if type_line[2] == 'NUMERIC':\n",
    "                if '\"^^' in value:\n",
    "                    value = value[1:value.index('\"^^')]\n",
    "                try:\n",
    "                    v[current] = str(float(value))\n",
    "                except ValueError:\n",
    "                    v[current] = '0.0'\n",
    "                current += 1\n",
    "            elif type_line[2] != 'IGNORE':\n",
    "                pos = None\n",
    "                if value == 'NA' or value not in cat_to_pos[idx]:\n",
    "                    pos = cat_to_pos[idx]['OTHER']\n",
    "                else:\n",
    "                    pos = cat_to_pos[idx][value]\n",
    "                for idx2 in range(len(type_line) - 2):\n",
    "                    if idx2 == pos:\n",
    "                        v[current] = \"1.0\"\n",
    "                    else:\n",
    "                        v[current] = \"0.0\"\n",
    "                    current += 1\n",
    "        feat1.write('\\t'.join(v) + \"\\n\")        \n",
    "print(\"Vector size\", total_size)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This data set is ready to be used with SVR. Doing a grid search for its parameters and saving data for error analysis (Cell #22) we obtain a RMSE of 0.41 and the curve below."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Training on 35,971 cities\n",
      "RMSE 0.41176199300264366\n",
      "Writing files for error analysis\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x7f92c6c90bd0>"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x360 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 22\n",
    "import random\n",
    "import math\n",
    "from sklearn.svm import SVR\n",
    "import numpy as np\n",
    "\n",
    "rand = random.Random(42)\n",
    "train_data = list()\n",
    "test_data = list()\n",
    "header = None\n",
    "with open(\"ch6_cell21_dev_feat1.tsv\") as feats:\n",
    "    header = next(feats)\n",
    "    header = header.strip().split(\"\\t\")\n",
    "    for line in feats:\n",
    "        fields = line.strip().split(\"\\t\")\n",
    "        pop = float(fields[-1])\n",
    "        name = fields[0]\n",
    "        feats = list(map(float,fields[1:-1]))\n",
    "        row = (feats, math.log(pop, 10), name)\n",
    "        if rand.random() < 0.2:\n",
    "            test_data.append(row) \n",
    "        else:\n",
    "            train_data.append(row)\n",
    "               \n",
    "test_data = sorted(test_data, key=lambda t:t[1])\n",
    "test_names = list(map(lambda t:t[2], test_data))\n",
    "\n",
    "xtrain = np.array(list(map(lambda t:t[0], train_data)))\n",
    "ytrain = np.array(list(map(lambda t:t[1], train_data)))\n",
    "xtest = np.array(list(map(lambda t:t[0], test_data)))\n",
    "ytest = np.array(list(map(lambda t:t[1], test_data)))\n",
    "\n",
    "# SVRs need scaling\n",
    "xtrain_min = xtrain.min(axis=0); xtrain_max = xtrain.max(axis=0)\n",
    "# some can be zero if the column is constant in training\n",
    "xtrain_diff = xtrain_max - xtrain_min\n",
    "for idx in range(len(xtrain_diff)):\n",
    "    if xtrain_diff[idx] == 0.0:\n",
    "        xtrain_diff[idx] = 1.0\n",
    "xtrain_scaling = 1.0 / xtrain_diff\n",
    "xtrain -= xtrain_min; xtrain *= xtrain_scaling\n",
    "\n",
    "ytrain_min = ytrain.min(); ytrain_max = ytrain.max()\n",
    "ytrain_scaling = 1.0 / (ytrain_max - ytrain_min)\n",
    "ytrain -= ytrain_min; ytrain *= ytrain_scaling\n",
    "\n",
    "xtest -= xtrain_min; xtest *= xtrain_scaling\n",
    "ytest_orig = ytest.copy()\n",
    "ytest -= ytrain_min; ytest *= ytrain_scaling\n",
    "\n",
    "# train\n",
    "print(\"Training on {:,} cities\".format(len(xtrain)))\n",
    "\n",
    "best_c = 100.0\n",
    "best_epsilon = 0.05\n",
    "\n",
    "PARAM_DO_GRID_SEARCH = False\n",
    "if PARAM_DO_GRID_SEARCH:\n",
    "    best_rmse = 1000\n",
    "    for c in [0.01, 0.1, 1.0, 2.0, 5.0, 10.0, 50.0, 100.0]:\n",
    "        svr_rbf = SVR(epsilon=0.05, C=c, gamma='auto')\n",
    "        svr_rbf.fit(xtrain, ytrain)\n",
    "        ytest_pred = svr_rbf.predict(xtest)\n",
    "        ytest_pred *= 1.0/ytrain_scaling\n",
    "        ytest_pred += ytrain_min\n",
    "        RMSE = math.sqrt(sum((ytest_orig - ytest_pred)**2) / len(ytest))\n",
    "        print(\"C\", c, \"RMSE\", RMSE)\n",
    "        if RMSE < best_rmse:\n",
    "            best_c = c\n",
    "            best_rmse = RMSE\n",
    "\n",
    "    print(\"Best C\", best_c,\"best RMSE\", best_rmse)        \n",
    "\n",
    "    best_rmse = 1000\n",
    "    for epsilon in [0.0001, 0.001, 0.01, 0.05, 0.1, 0.5, 0.0, 1.0, 10.0, 100.0]:\n",
    "        svr_rbf = SVR(epsilon=epsilon, C=best_c, gamma='auto')\n",
    "        svr_rbf.fit(xtrain, ytrain)\n",
    "        ytest_pred = svr_rbf.predict(xtest)\n",
    "        ytest_pred *= 1.0/ytrain_scaling\n",
    "        ytest_pred += ytrain_min\n",
    "        RMSE = math.sqrt(sum((ytest_orig - ytest_pred)**2) / len(ytest))\n",
    "        print(\"Epsilon\", epsilon, \"RMSE\", RMSE)\n",
    "        if RMSE < best_rmse:\n",
    "            best_epsilon = epsilon\n",
    "            best_rmse = RMSE\n",
    "\n",
    "    print(\"Best epsilon\", best_epsilon,\"best RMSE\", best_rmse)\n",
    "svr_rbf = SVR(epsilon=best_epsilon, C=best_c, gamma='auto')\n",
    "svr_rbf.fit(xtrain, ytrain)\n",
    "ytest_pred = svr_rbf.predict(xtest)\n",
    "ytest_pred *= 1.0/ytrain_scaling\n",
    "ytest_pred += ytrain_min\n",
    "RMSE = math.sqrt(sum((ytest_orig - ytest_pred)**2) / len(ytest))\n",
    "print(\"RMSE\", RMSE)\n",
    "\n",
    "print(\"Writing files for error analysis\")\n",
    "np.savez_compressed(\"ch6_cell22_feat1_svr.npz\", \n",
    "                    xtrain=xtrain, ytrain=ytrain, \n",
    "                    xtest=xtest, ytest=ytest,\n",
    "                    ytest_orig=ytest_orig,\n",
    "                    ytest_pred=ytest_pred,\n",
    "                    ytrain_min=ytrain_min,\n",
    "                    xtrain_scaling=xtrain_scaling,\n",
    "                    ytrain_scaling=ytrain_scaling)\n",
    "with open(\"ch6_cell22_test_names.tsv\", \"w\") as names:\n",
    "    for idx, name in enumerate(test_names):\n",
    "        names.write(\"{}\\t{}\\n\".format(idx, name))\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "plt.rcParams['figure.figsize'] = [20, 5]\n",
    "plt.plot(ytest_pred, label=\"predicted\", color='gray')\n",
    "plt.plot(ytest_orig, label=\"actual\",    color='black')\n",
    "plt.ylabel('scaled log population')\n",
    "plt.savefig(\"ch6_cell22_svr_feat1.pdf\", bbox_inches='tight', dpi=300)\n",
    "plt.legend()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This figure and RMSE look better than the single feature that preceed it but it has some strange spikes. Let's perform some Error Analysis on them, an error drilldown on the spike at the end by feature ablation, most probably won't work but I show error analysis.\n",
    "\n",
    "### Error Analysis\n",
    "\n",
    "Using the training data and evaluation results from Cell #22, I can now do an error drilldown on the spike at the end by feature ablation in Cell #23."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Top contributors to the error\n",
      "8825  Dublin                  6.5       553,164   1,751,459,207,216\n",
      "8726  Saint-Prex            -2.09       312,010               2,485\n",
      "8705  Meads,_Kentucky       -2.08       288,648               2,351\n",
      "7939  Nicosia                2.03        55,013           5,931,549\n",
      "8921  Hofuf                 -1.94     1,500,000              17,145\n",
      "8985  Mexico_City            1.93     8,918,652         761,989,297\n",
      "8689  Montana_City,_Montana -1.88       271,529               3,513\n",
      "8902  Pikine                -1.82     1,170,790              17,470\n",
      "8791  Edinburgh              1.76       464,989          26,976,492\n",
      "172   Isabela,_Basilan       1.74         1,081              60,349\n"
     ]
    }
   ],
   "source": [
    "# CELL 23\n",
    "arrays = np.load(\"ch6_cell22_feat1_svr.npz\")\n",
    "ytest_orig = arrays['ytest_orig']\n",
    "ytest_pred = arrays['ytest_pred']\n",
    "test_names = list(map(lambda line:line.strip().split('\\t')[-1], \n",
    "                      open(\"ch6_cell22_test_names.tsv\").readlines()))\n",
    "\n",
    "# find the top 10 biggest contributors to the log error\n",
    "sqe = (ytest_orig - ytest_pred)**2\n",
    "contributors = list()\n",
    "while len(contributors) < 10:\n",
    "    idx = sqe.argmax()\n",
    "    contributors.append( (idx, (ytest_pred - ytest_orig)[idx]) )\n",
    "    sqe[idx] = 0.0\n",
    "\n",
    "print(\"Top contributors to the error\")\n",
    "for idx, error in contributors:\n",
    "    print(\"{:<4}  {:<21} {:>5}    {:>10,}   {:>17,}\".\n",
    "          format(idx, test_names[idx][:-1].split(\"/\")[-1], int(error*100)/100, \n",
    "          int(10**ytest_orig[idx]), int(10**ytest_pred[idx])))                 "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "From the table we can see that cities that have a rich history but are not necessarily that populated get overshoot by the system. Makes sense.\n",
    "\n",
    "Interestingly, Saint-Prex is actually about 5,000 people, the 300,000 number is an extraction mistake. Meads, Kentucky population is listed on Wikipedia as 280,000 but it is being flagged as an error as 280,000 is the size of the full region. The town itself is definitely less than 5,000 people.\n",
    "\n",
    "Nicosia is a history-rich city populated by 4,500 years. Sadly, it is a divided city and one of its divisions population is about 55,000 people. A joint city will total about 300,000 inhabitants, but it has the history of a 6,000,000 people for sure.\n",
    "\n",
    "Therefore we have history-rich, underpopulated cities and extraction errors. From here it makes sense to focus on real (not logarithmic) where the system is over predicting (Cell #24)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Top contributors to the error\n",
      "Cumulative error 17.634868409695972\n",
      "8825  Dublin                  6.5       553,164   1,751,459,207,216\n",
      "8985  Mexico_City            1.93     8,918,652         761,989,297\n",
      "8791  Edinburgh              1.76       464,989          26,976,492\n",
      "8959  Algiers                0.83     3,415,810          23,294,353\n",
      "8910  Prague                 0.88     1,267,448           9,823,330\n",
      "8915  Milan                   0.8     1,359,904           8,636,483\n",
      "8873  Amsterdam              0.98       842,342           8,094,965\n",
      "8823  Lisbon                 1.15       545,245           7,776,212\n",
      "8924  Tabriz                 0.74     1,549,452           8,586,162\n",
      "7939  Nicosia                2.03        55,013           5,931,549\n"
     ]
    }
   ],
   "source": [
    "# CELL 24\n",
    "arrays = np.load(\"ch6_cell22_feat1_svr.npz\")\n",
    "xtrain = arrays['xtrain']\n",
    "ytrain = arrays['ytrain'] \n",
    "xtest = arrays['xtest']\n",
    "ytest = arrays['ytest']\n",
    "ytest_orig = arrays['ytest_orig']\n",
    "ytest_pred = arrays['ytest_pred']\n",
    "test_names = list(map(lambda line:line.strip().split('\\t')[-1], \n",
    "                      open(\"ch6_cell22_test_names.tsv\").readlines()))\n",
    "\n",
    "# find the top 10 biggest contributors to the actual error\n",
    "real_error = (10**ytest_pred - 10**ytest_orig)\n",
    "contributors = list()\n",
    "cumulative_error = 0\n",
    "while len(contributors) < 10:\n",
    "    idx = real_error.argmax()\n",
    "    contributors.append( (idx, (ytest_pred - ytest_orig)[idx]) )\n",
    "    cumulative_error += ytest_pred[idx] - ytest_orig[idx]\n",
    "    real_error[idx] = 0.0\n",
    "\n",
    "print(\"Top contributors to the error\")\n",
    "print(\"Cumulative error {:,}\".format(cumulative_error))\n",
    "for idx, error in contributors:\n",
    "    print(\"{:<4}  {:<21} {:>5}    {:>10,}   {:>17,}\".\n",
    "          format(idx, test_names[idx][:-1].split(\"/\")[-1], int(error*100)/100, \n",
    "          int(10**ytest_orig[idx]), int(10**ytest_pred[idx])))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This list looks quite good, these cities are contributing a lot to the error and fall into the category of errors we have found the system is making, let's do an ablation study (wrapper) to see if we can find which features are hurting us here. In this study we'll throw away one feature at a time and see whether it improves the cumulative error for these 10 cities. The amount of computation is very big so we'll work on a 1% sample of the original data. Also, because Dublin is so markedly overshoot, I'll also keep track of dropping which better feature improves each city. Cell #25."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "  0 out of 380\n",
      " 38 out of 380\n",
      " 76 out of 380\n",
      "114 out of 380\n",
      "152 out of 380\n",
      "190 out of 380\n",
      "228 out of 380\n",
      "266 out of 380\n",
      "304 out of 380\n",
      "342 out of 380\n",
      "Ablation study\n",
      "Best overall from 0.28 to -2.4 (968.43%)\n",
      "Feature to ablate (303): http://www.w3.org/2000/01/rdf-schema#seeAlso#count\n",
      "\n",
      "\n",
      "For city Dublin from 1.2 to 0.48 (58.97%)\n",
      "Feature to ablate (11): http://dbpedia.org/ontology/city?inv#count\n",
      "\n",
      "For city Mexico_City from -0.061 to -0.4 (-561.55%)\n",
      "Feature to ablate (303): http://www.w3.org/2000/01/rdf-schema#seeAlso#count\n",
      "\n",
      "For city Edinburgh from 0.43 to 0.19 (57.03%)\n",
      "Feature to ablate (3): http://dbpedia.org/ontology/areaLand#count\n",
      "\n",
      "For city Algiers from -0.92 to -1.1 (-22.87%)\n",
      "Feature to ablate (378): http://xmlns.com/foaf/0.1/name#count\n",
      "\n",
      "For city Prague from -0.18 to -0.82 (-341.93%)\n",
      "Feature to ablate (303): http://www.w3.org/2000/01/rdf-schema#seeAlso#count\n",
      "\n",
      "For city Milan from 0.17 to -0.43 (347.88%)\n",
      "Feature to ablate (307): http://www.w3.org/2000/01/rdf-schema#seeAlso#3@<http://dbpedia.org/resource/Italy>\n",
      "\n",
      "For city Amsterdam from -0.53 to -0.7 (-31.85%)\n",
      "Feature to ablate (98): http://dbpedia.org/ontology/ground?inv#count\n",
      "\n",
      "For city Lisbon from 0.0019 to -0.16 (8666.74%)\n",
      "Feature to ablate (323): http://www.w3.org/2000/01/rdf-schema#seeAlso#3@<http://dbpedia.org/resource/The_Republic_of_Ireland>\n",
      "\n",
      "For city Tabriz from -0.2 to -0.97 (-376.48%)\n",
      "Feature to ablate (303): http://www.w3.org/2000/01/rdf-schema#seeAlso#count\n",
      "\n",
      "For city Nicosia from 0.41 to 0.1 (75.64%)\n",
      "Feature to ablate (303): http://www.w3.org/2000/01/rdf-schema#seeAlso#count\n",
      "\n"
     ]
    }
   ],
   "source": [
    "# CELL 25\n",
    "import random\n",
    "from sklearn.svm import SVR\n",
    "import numpy as np\n",
    "\n",
    "arrays = np.load(\"ch6_cell22_feat1_svr.npz\")\n",
    "xtrain = arrays['xtrain']\n",
    "ytrain = arrays['ytrain']\n",
    "ytrain_scaling = arrays['ytrain_scaling']\n",
    "ytrain_min = arrays['ytrain_min']\n",
    "xtest = arrays['xtest']\n",
    "ytest = arrays['ytest']\n",
    "ytest_orig = arrays['ytest_orig']\n",
    "ytest_pred = arrays['ytest_pred']\n",
    "test_names = list(map(lambda line:line.strip().split('\\t')[-1], \n",
    "                      open(\"ch6_cell22_test_names.tsv\").readlines()))\n",
    "header = None\n",
    "with open(\"ch6_cell21_dev_feat1.tsv\") as feats:\n",
    "    header = next(feats)\n",
    "header = header.strip().split('\\t')\n",
    "header.pop(0) # drop name\n",
    "\n",
    "# find the top 10 biggest contributors to the log error\n",
    "real_error = (10**ytest_pred - 10**ytest_orig)\n",
    "contributors = list()\n",
    "cumulative_error = 0\n",
    "xtest_contr = np.zeros( (10, xtest.shape[1]) )\n",
    "ytest_contr = np.zeros( (10,) )\n",
    "while len(contributors) < 10:\n",
    "    idx = real_error.argmax()\n",
    "    ytest_contr[len(contributors)] = ytest_orig[idx]\n",
    "    xtest_contr[len(contributors),:] = xtest[idx,:]\n",
    "    contributors.append( (idx, (ytest_pred - ytest_orig)[idx]) )\n",
    "    cumulative_error += real_error[idx]\n",
    "    real_error[idx] = 0.0\n",
    "\n",
    "best_epsilon = 0.05\n",
    "best_c = 100.0\n",
    "# this is a lot of computation, work on sample of the original data\n",
    "PARAM_SAMPLE_PERC=0.05\n",
    "rand = random.Random(42)\n",
    "selected = rand.sample(list(range(len(xtrain))), int(len(xtrain) * PARAM_SAMPLE_PERC))\n",
    "xtrain_samp = xtrain[selected,:]\n",
    "ytrain_samp = ytrain[selected]\n",
    "\n",
    "# baseline run\n",
    "svr_rbf = SVR(epsilon=best_epsilon, C=best_c, gamma='auto')\n",
    "svr_rbf.fit(xtrain_samp, ytrain_samp)\n",
    "ytest_pred = svr_rbf.predict(xtest_contr)\n",
    "ytest_pred *= 1.0/ytrain_scaling\n",
    "ytest_pred += ytrain_min\n",
    "error = ytest_pred - ytest_contr # absolute log error\n",
    "cumulative = sum(error)\n",
    "baseline_cumulative = cumulative\n",
    "per_city_baseline = error\n",
    "\n",
    "best_cumulative = 100000.00\n",
    "best_cumulative_ablate = None\n",
    "per_city_best = [10000.00,] * 10\n",
    "per_city_best_ablate = [None,] * 10\n",
    "\n",
    "for feat in range(xtrain.shape[1]):\n",
    "    if feat % int(xtrain.shape[1] / 10) == 0:\n",
    "        print(\"{:>3} out of {:>3}\".format(feat, xtrain.shape[1]))\n",
    "    train_col = xtrain_samp[:,feat].copy()\n",
    "    xtrain_samp[:,feat] = 0.0 # nuke feature\n",
    "    svr_rbf = SVR(epsilon=best_epsilon, C=best_c, gamma='auto')\n",
    "    svr_rbf.fit(xtrain_samp, ytrain_samp)\n",
    "    ytest_pred = svr_rbf.predict(xtest_contr)\n",
    "    ytest_pred *= 1.0/ytrain_scaling\n",
    "    ytest_pred += ytrain_min\n",
    "    error = ytest_pred - ytest_contr\n",
    "    cumulative = sum(error)\n",
    "    if best_cumulative_ablate is None or best_cumulative > cumulative:\n",
    "        best_cumulative_ablate = feat\n",
    "        best_cumulative = cumulative\n",
    "    for idx in range(10):\n",
    "        if per_city_best_ablate[idx] is None or per_city_best[idx] > error[idx]:\n",
    "            per_city_best_ablate[idx] = feat\n",
    "            per_city_best[idx] = error[idx]\n",
    "    xtrain_samp[:,feat] = train_col # restore\n",
    "\n",
    "print(\"Ablation study\")\n",
    "print(\"Best overall from {:3.2} to {:3.2} ({:.2%})\".format(\n",
    "    baseline_cumulative, best_cumulative,\n",
    "    (baseline_cumulative-best_cumulative)/baseline_cumulative))\n",
    "print(\"Feature to ablate ({}): {}\\n\\n\".format(best_cumulative_ablate, header[best_cumulative_ablate]))\n",
    "for idx in range(10):\n",
    "    print(\"For city {} from {:3.2} to {:3.2} ({:.2%})\".format(\n",
    "        test_names[contributors[idx][0]][:-1].split(\"/\")[-1],\n",
    "        per_city_baseline[idx], per_city_best[idx],\n",
    "        (per_city_baseline[idx]-per_city_best[idx])/per_city_baseline[idx]))\n",
    "    print(\"Feature to ablate ({}): {}\\n\".format(per_city_best_ablate[idx], header[per_city_best_ablate[idx]]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "From the above results I can see that counts are misleading the algorithm. Here I am presented with two options, I can either drop them or I can dampen their growth by applying a squashing function to them. As I believe there's value in them, I'll dampen their value by applying a log function to them (Cell #26)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Training on 35,971 cities\n",
      "RMSE 0.3556252788046216\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x7f9287312f90>"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x360 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 26\n",
    "import math\n",
    "from sklearn.svm import SVR\n",
    "import numpy as np\n",
    "\n",
    "header = None\n",
    "with open(\"ch6_cell26_dev_feat1_dampen.tsv\", \"w\") as feats_damp:\n",
    "    with open(\"ch6_cell21_dev_feat1.tsv\") as feats:\n",
    "        header = next(feats)\n",
    "        feats_damp.write(header)\n",
    "        header = header.strip().split(\"\\t\")\n",
    "        header.pop(0) # name\n",
    "        for line in feats:\n",
    "            fields = line.strip().split(\"\\t\")\n",
    "            pop = float(fields[-1])\n",
    "            name = fields[0]\n",
    "            feats = list(map(float,fields[1:-1]))\n",
    "            pop = math.log(pop, 10)\n",
    "            # dampen counts\n",
    "            for idx, featname in enumerate(header):\n",
    "                if \"#count\" in featname:\n",
    "                    feats[idx] = math.log(feats[idx] + 1, 10)\n",
    "            feats_damp.write(\"\\t\".join(map(str, [name] + feats + [pop])) + \"\\n\")\n",
    "\n",
    "rand = random.Random(42)\n",
    "train_data = list()\n",
    "test_data = list()\n",
    "header = None\n",
    "with open(\"ch6_cell26_dev_feat1_dampen.tsv\") as feats:\n",
    "    header = next(feats)\n",
    "    header = header.strip().split(\"\\t\")\n",
    "    header.pop(0) # name\n",
    "    for line in feats:\n",
    "        fields = line.strip().split(\"\\t\")\n",
    "        logpop = float(fields[-1])\n",
    "        name = fields[0]\n",
    "        feats = list(map(float,fields[1:-1]))\n",
    "        row = (feats, logpop, name)\n",
    "        if rand.random() < 0.2:\n",
    "            test_data.append(row) \n",
    "        else:\n",
    "            train_data.append(row)\n",
    "               \n",
    "test_data = sorted(test_data, key=lambda t:t[1])\n",
    "test_names = list(map(lambda t:t[2], test_data))\n",
    "\n",
    "xtrain = np.array(list(map(lambda t:t[0], train_data)))\n",
    "ytrain = np.array(list(map(lambda t:t[1], train_data)))\n",
    "xtest = np.array(list(map(lambda t:t[0], test_data)))\n",
    "ytest = np.array(list(map(lambda t:t[1], test_data)))\n",
    "\n",
    "# SVRs need scaling\n",
    "xtrain_min = xtrain.min(axis=0); xtrain_max = xtrain.max(axis=0)\n",
    "# some can be zero if the column is constant in training\n",
    "xtrain_diff = xtrain_max - xtrain_min\n",
    "for idx in range(len(xtrain_diff)):\n",
    "    if xtrain_diff[idx] == 0.0:\n",
    "        xtrain_diff[idx] = 1.0\n",
    "xtrain_scaling = 1.0 / xtrain_diff\n",
    "xtrain -= xtrain_min; xtrain *= xtrain_scaling\n",
    "\n",
    "ytrain_min = ytrain.min(); ytrain_max = ytrain.max()\n",
    "ytrain_scaling = 1.0 / (ytrain_max - ytrain_min)\n",
    "ytrain -= ytrain_min; ytrain *= ytrain_scaling\n",
    "\n",
    "xtest -= xtrain_min; xtest *= xtrain_scaling\n",
    "ytest_orig = ytest.copy()\n",
    "ytest -= ytrain_min; ytest *= ytrain_scaling\n",
    "\n",
    "# train\n",
    "print(\"Training on {:,} cities\".format(len(xtrain)))\n",
    "\n",
    "best_c = 100.0\n",
    "best_epsilon = 0.05\n",
    "svr_rbf = SVR(epsilon=best_epsilon, C=best_c, gamma='auto')\n",
    "svr_rbf.fit(xtrain, ytrain)\n",
    "ytest_pred = svr_rbf.predict(xtest)\n",
    "ytest_pred *= 1.0/ytrain_scaling\n",
    "ytest_pred += ytrain_min\n",
    "RMSE = math.sqrt(sum((ytest_orig - ytest_pred)**2) / len(ytest))\n",
    "print(\"RMSE\", RMSE)\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "plt.rcParams['figure.figsize'] = [20, 5]\n",
    "plt.plot(ytest_pred, label=\"predicted\", color='gray')\n",
    "plt.plot(ytest_orig, label=\"actual\", color='black')\n",
    "plt.ylabel('scaled log population')\n",
    "plt.savefig(\"ch6_cell26_svr_feat1_dampen.pdf\", bbox_inches='tight', dpi=300)\n",
    "plt.legend()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The RMSE is better than before at 0.36 and the spikes have reduced.\n",
    "\n",
    "That is a nice improvement from the previous one (note that Feature Engineering does not always succeed and you might encounter dead ends). (Of course taking the log of those counts could have been done directly based on the advice in Chapter 2).\n",
    "\n",
    "Instead of continuing in this route of wrapper ablation, I'll now move to feature utility metrics, for which we need a categorical target, I'll thus discretize the target in segments with equal number of points. Cell #27 does the discretization and measure the discretization error at different number of splits."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\tSegment  1, min:      1,000 value:      5,409, max: 24,300,000\n",
      "Segmented in 1, RMSE: 0.703382635244624\n",
      "\tSegment  1, min:      1,000 value:      2,226, max:      5,409\n",
      "\tSegment  2, min:      5,410 value:     19,435, max: 24,300,000\n",
      "Segmented in 2, RMSE: 0.44244189595902794\n",
      "\tSegment  1, min:      1,000 value:      1,545, max:      2,226\n",
      "\tSegment  2, min:      2,226 value:      3,363, max:      5,409\n",
      "\tSegment  3, min:      5,410 value:      9,734, max:     19,435\n",
      "\tSegment  4, min:     19,435 value:     49,331, max: 24,300,000\n",
      "Segmented in 4, RMSE: 0.2912613798706353\n",
      "\tSegment  1, min:      1,000 value:      1,254, max:      1,545\n",
      "\tSegment  2, min:      1,545 value:      1,884, max:      2,226\n",
      "\tSegment  3, min:      2,226 value:      2,729, max:      3,363\n",
      "\tSegment  4, min:      3,364 value:      4,225, max:      5,409\n",
      "\tSegment  5, min:      5,410 value:      7,147, max:      9,734\n",
      "\tSegment  6, min:      9,735 value:     13,515, max:     19,435\n",
      "\tSegment  7, min:     19,435 value:     29,261, max:     49,339\n",
      "\tSegment  8, min:     49,371 value:    114,914, max: 24,300,000\n",
      "Segmented in 8, RMSE: 0.1936688721250952\n",
      "\tSegment  1, min:      1,000 value:      1,126, max:      1,254\n",
      "\tSegment  2, min:      1,254 value:      1,393, max:      1,545\n",
      "\tSegment  3, min:      1,545 value:      1,710, max:      1,884\n",
      "\tSegment  4, min:      1,884 value:      2,028, max:      2,226\n",
      "\tSegment  5, min:      2,226 value:      2,461, max:      2,729\n",
      "\tSegment  6, min:      2,729 value:      3,021, max:      3,363\n",
      "\tSegment  7, min:      3,364 value:      3,766, max:      4,225\n",
      "\tSegment  8, min:      4,225 value:      4,800, max:      5,409\n",
      "\tSegment  9, min:      5,410 value:      6,176, max:      7,147\n",
      "\tSegment 10, min:      7,147 value:      8,300, max:      9,734\n",
      "\tSegment 11, min:      9,735 value:     11,441, max:     13,518\n",
      "\tSegment 12, min:     13,518 value:     16,117, max:     19,435\n",
      "\tSegment 13, min:     19,435 value:     23,699, max:     29,261\n",
      "\tSegment 14, min:     29,271 value:     37,076, max:     49,339\n",
      "\tSegment 15, min:     49,371 value:     69,696, max:    114,914\n",
      "\tSegment 16, min:    114,924 value:    264,091, max: 24,300,000\n",
      "Segmented in 16, RMSE: 0.12719110586720506\n",
      "\tSegment  1, min:      1,000 value:      1,068, max:      1,126\n",
      "\tSegment  2, min:      1,126 value:      1,189, max:      1,254\n",
      "\tSegment  3, min:      1,254 value:      1,318, max:      1,393\n",
      "\tSegment  4, min:      1,393 value:      1,461, max:      1,545\n",
      "\tSegment  5, min:      1,545 value:      1,628, max:      1,710\n",
      "\tSegment  6, min:      1,710 value:      1,799, max:      1,884\n",
      "\tSegment  7, min:      1,884 value:      1,986, max:      2,028\n",
      "\tSegment  8, min:      2,028 value:      2,125, max:      2,226\n",
      "\tSegment  9, min:      2,226 value:      2,343, max:      2,461\n",
      "\tSegment 10, min:      2,461 value:      2,588, max:      2,729\n",
      "\tSegment 11, min:      2,729 value:      2,883, max:      3,021\n",
      "\tSegment 12, min:      3,022 value:      3,175, max:      3,363\n",
      "\tSegment 13, min:      3,364 value:      3,566, max:      3,766\n",
      "\tSegment 14, min:      3,766 value:      3,995, max:      4,225\n",
      "\tSegment 15, min:      4,225 value:      4,503, max:      4,800\n",
      "\tSegment 16, min:      4,800 value:      5,098, max:      5,409\n",
      "\tSegment 17, min:      5,410 value:      5,790, max:      6,176\n",
      "\tSegment 18, min:      6,177 value:      6,609, max:      7,147\n",
      "\tSegment 19, min:      7,147 value:      7,720, max:      8,300\n",
      "\tSegment 20, min:      8,300 value:      8,963, max:      9,734\n",
      "\tSegment 21, min:      9,735 value:     10,546, max:     11,441\n",
      "\tSegment 22, min:     11,442 value:     12,419, max:     13,518\n",
      "\tSegment 23, min:     13,518 value:     14,776, max:     16,117\n",
      "\tSegment 24, min:     16,119 value:     17,742, max:     19,435\n",
      "\tSegment 25, min:     19,435 value:     21,342, max:     23,699\n",
      "\tSegment 26, min:     23,700 value:     26,171, max:     29,261\n",
      "\tSegment 27, min:     29,271 value:     32,882, max:     37,080\n",
      "\tSegment 28, min:     37,095 value:     42,033, max:     49,339\n",
      "\tSegment 29, min:     49,371 value:     57,331, max:     69,713\n",
      "\tSegment 30, min:     69,769 value:     87,249, max:    114,914\n",
      "\tSegment 31, min:    114,924 value:    161,719, max:    264,700\n",
      "\tSegment 32, min:    264,716 value:    655,965, max: 24,300,000\n",
      "Segmented in 32, RMSE: 0.07988951417940254\n"
     ]
    }
   ],
   "source": [
    "# CELL 27\n",
    "import math\n",
    "import random\n",
    "import numpy as np\n",
    "\n",
    "rand = random.Random(42)\n",
    "train_data = list()\n",
    "test_data = list()\n",
    "with open(\"ch6_cell26_dev_feat1_dampen.tsv\") as feats:\n",
    "    header = next(feats)\n",
    "    for line in feats:\n",
    "        fields = line.strip().split(\"\\t\")\n",
    "        logpop = float(fields[-1])\n",
    "        if rand.random() < 0.2:\n",
    "            test_data.append(logpop) \n",
    "        else:\n",
    "            train_data.append(logpop)\n",
    "\n",
    "train_data = np.array(sorted(train_data))\n",
    "test_data  = np.array(test_data)\n",
    "\n",
    "def cell27_split(data):\n",
    "    pivot = int(len(data)/2)\n",
    "    # slower, more accurate\n",
    "    lower_center, lower_val = int(pivot/2), data[int(pivot/2)]\n",
    "    upper_center, upper_val = int(pivot/2)+pivot, data[int(pivot/2)+pivot]\n",
    "    return [ { \"min\": data[0], \"max\": data[pivot], \"val\": lower_val,\n",
    "              \"min_idx\": 0, \"max_idx\": pivot, \"center\": lower_center },\n",
    "             { \"min\": data[pivot+1], \"max\": data[-1], \"val\": upper_val,\n",
    "              \"min_idx\": pivot+1, \"max_idx\": len(data) - 1, \"center\": upper_center } ]\n",
    "\n",
    "def cell27_segment(train_data, order):\n",
    "    segments_at = [ [ { \"min\": train_data[0], \"max\" : train_data[-1], \n",
    "                        \"val\": train_data[int(len(train_data)/2)],\n",
    "                        \"min_idx\":0, \"max_idx\": len(train_data) - 1 } ] ]\n",
    "    for exp in range(order):\n",
    "        previous = segments_at[-1]\n",
    "        new_segment = list()\n",
    "        for old_segment in previous:\n",
    "            splitted = cell27_split(train_data[old_segment['min_idx']:(old_segment['max_idx']+1)])\n",
    "            for created in splitted:\n",
    "                created['min_idx'] += old_segment['min_idx']\n",
    "                created['max_idx'] += old_segment['min_idx']\n",
    "                created['center']  += old_segment['min_idx']\n",
    "                new_segment.append(created)\n",
    "        segments_at.append(new_segment)\n",
    "    return segments_at\n",
    "\n",
    "def cell27_predict(data, segments):\n",
    "    result = list()\n",
    "    for pop in data:\n",
    "        value = None\n",
    "        if pop < segments[0]['min']:\n",
    "            value = segments[0]['val']\n",
    "        elif pop > segments[-1]['max']:\n",
    "            value = segments[-1]['val']\n",
    "        else:\n",
    "            for seg_idx, segment in enumerate(segments):\n",
    "                if segment['min'] <= pop and \\\n",
    "                        (seg_idx == len(segments) - 1 or pop < segments[seg_idx + 1]['min']):\n",
    "                    value = segment['val']\n",
    "                    break\n",
    "        if value is None:\n",
    "            raise Exception(\"Value not found: {}\".format(10**pop))\n",
    "        result.append(value)\n",
    "    return np.array(result)\n",
    "\n",
    "segments_at = cell27_segment(train_data, 5)\n",
    "\n",
    "for idx, segments in enumerate(segments_at):\n",
    "    for seg_idx, seg in enumerate(segments):\n",
    "        print(\"\\tSegment {:2}, min: {:10,} value: {:10,}, max: {:10,}\"\n",
    "              .format((seg_idx+1), int(round(10 ** seg['min'])), \n",
    "                      int(round(10 ** seg['val'])), int(round(10 ** seg['max']))))\n",
    "    test_pred = cell27_predict(test_data, segments)\n",
    "    print(\"Segmented in {}, RMSE: {}\".format(2**idx, math.sqrt(sum((test_data - test_pred)**2)/len(test_data))))\n",
    "    \n",
    "import pickle\n",
    "with open(\"ch6_cell27_splits.pk\", \"wb\") as pkl:\n",
    "    pickle.dump(segments_at, pkl)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "From above, splitting into 4 segments produce a reasonable discretiation error. I can now proceed to do some feature selection using mutual information (Chapter 4). I will also introduce 10 random features and drop all features whose mutual information falls below at least 3 random features. Cell #28."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "  0 out of 390\n",
      " 39 out of 390\n",
      " 78 out of 390\n",
      "117 out of 390\n",
      "156 out of 390\n",
      "195 out of 390\n",
      "234 out of 390\n",
      "273 out of 390\n",
      "312 out of 390\n",
      "351 out of 390\n",
      "Dropped 68 features\n",
      "New feature vector size 314 features, including name and target\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<h3>Top 150 features by MI</h3><table><tr><th>Position</th><th>Feat#</th><th>Feat</th><th>Utility</th></tr><tr><td>0</td><td>286</td><td>http://dbpedia.org/ontology/utcOffset#count</td><td>0.0976963935</td></tr>\n",
       "<tr><td>1</td><td>10</td><td>http://dbpedia.org/ontology/birthPlace?inv#count</td><td>0.0773520859</td></tr>\n",
       "<tr><td>2</td><td>33</td><td>http://dbpedia.org/ontology/country#1@OTHER</td><td>0.0715361164</td></tr>\n",
       "<tr><td>3</td><td>7</td><td>http://dbpedia.org/ontology/areaTotal#1</td><td>0.0574203739</td></tr>\n",
       "<tr><td>4</td><td>301</td><td>http://www.w3.org/1999/02/22-rdf-syntax-ns#type#1@http://dbpedia.org/ontology/City</td><td>0.0446595875</td></tr>\n",
       "<tr><td>5</td><td>100</td><td>http://dbpedia.org/ontology/hometown?inv#count</td><td>0.0444925574</td></tr>\n",
       "<tr><td>6</td><td>104</td><td>http://dbpedia.org/ontology/leaderTitle#count</td><td>0.0418022897</td></tr>\n",
       "<tr><td>7</td><td>98</td><td>http://dbpedia.org/ontology/ground?inv#count</td><td>0.0417288633</td></tr>\n",
       "<tr><td>8</td><td>102</td><td>http://dbpedia.org/ontology/isPartOf?inv#count</td><td>0.0383227077</td></tr>\n",
       "<tr><td>9</td><td>99</td><td>http://dbpedia.org/ontology/headquarter?inv#count</td><td>0.0381511672</td></tr>\n",
       "<tr><td>10</td><td>11</td><td>http://dbpedia.org/ontology/city?inv#count</td><td>0.0375958312</td></tr>\n",
       "<tr><td>11</td><td>110</td><td>http://dbpedia.org/ontology/locationCity?inv#count</td><td>0.0345756327</td></tr>\n",
       "<tr><td>12</td><td>109</td><td>http://dbpedia.org/ontology/location?inv#count</td><td>0.0344053176</td></tr>\n",
       "<tr><td>13</td><td>303</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#count</td><td>0.0329376538</td></tr>\n",
       "<tr><td>14</td><td>101</td><td>http://dbpedia.org/ontology/isPartOf#count</td><td>0.0310347434</td></tr>\n",
       "<tr><td>15</td><td>0</td><td>rel#count</td><td>0.0305265290</td></tr>\n",
       "<tr><td>16</td><td>103</td><td>http://dbpedia.org/ontology/leaderName#count</td><td>0.0300089032</td></tr>\n",
       "<tr><td>17</td><td>237</td><td>http://dbpedia.org/ontology/residence?inv#count</td><td>0.0297368517</td></tr>\n",
       "<tr><td>18</td><td>291</td><td>http://dbpedia.org/ontology/utcOffset#1@\"+8\"</td><td>0.0296566141</td></tr>\n",
       "<tr><td>19</td><td>379</td><td>http://xmlns.com/foaf/0.1/nick#count</td><td>0.0294899528</td></tr>\n",
       "<tr><td>20</td><td>378</td><td>http://xmlns.com/foaf/0.1/name#count</td><td>0.0294766008</td></tr>\n",
       "<tr><td>21</td><td>34</td><td>http://dbpedia.org/ontology/deathPlace?inv#count</td><td>0.0290410902</td></tr>\n",
       "<tr><td>22</td><td>289</td><td>http://dbpedia.org/ontology/utcOffset#1@\"+5:30\"</td><td>0.0258797043</td></tr>\n",
       "<tr><td>23</td><td>246</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/Indian_Standard_Time</td><td>0.0254364934</td></tr>\n",
       "<tr><td>24</td><td>17</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/India</td><td>0.0245425720</td></tr>\n",
       "<tr><td>25</td><td>299</td><td>http://dbpedia.org/ontology/utcOffset#2@OTHER</td><td>0.0223334458</td></tr>\n",
       "<tr><td>26</td><td>300</td><td>http://www.w3.org/1999/02/22-rdf-syntax-ns#type#1@http://dbpedia.org/ontology/Settlement</td><td>0.0219254334</td></tr>\n",
       "<tr><td>27</td><td>105</td><td>http://dbpedia.org/ontology/leaderTitle#1@\"Mayor\"@en</td><td>0.0211852222</td></tr>\n",
       "<tr><td>28</td><td>241</td><td>http://dbpedia.org/ontology/timeZone#count</td><td>0.0186070090</td></tr>\n",
       "<tr><td>29</td><td>13</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/United_States</td><td>0.0175165178</td></tr>\n",
       "<tr><td>30</td><td>253</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/China_Standard_Time</td><td>0.0170512916</td></tr>\n",
       "<tr><td>31</td><td>29</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Switzerland</td><td>0.0158900843</td></tr>\n",
       "<tr><td>32</td><td>94</td><td>http://dbpedia.org/ontology/district#count</td><td>0.0157178626</td></tr>\n",
       "<tr><td>33</td><td>97</td><td>http://dbpedia.org/ontology/foundingDate#count</td><td>0.0148247411</td></tr>\n",
       "<tr><td>34</td><td>267</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Cities_of_Japan</td><td>0.0141595020</td></tr>\n",
       "<tr><td>35</td><td>268</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Municipalities_of_Switzerland</td><td>0.0136140699</td></tr>\n",
       "<tr><td>36</td><td>260</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Communes_of_Romania</td><td>0.0134756551</td></tr>\n",
       "<tr><td>37</td><td>14</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/France</td><td>0.0119833788</td></tr>\n",
       "<tr><td>38</td><td>377</td><td>http://xmlns.com/foaf/0.1/homepage#count</td><td>0.0117127918</td></tr>\n",
       "<tr><td>39</td><td>30</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Brazil</td><td>0.0114840226</td></tr>\n",
       "<tr><td>40</td><td>19</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Romania</td><td>0.0109216127</td></tr>\n",
       "<tr><td>41</td><td>262</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Municipalities_of_the_Philippines</td><td>0.0108070888</td></tr>\n",
       "<tr><td>42</td><td>249</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/Philippine_Standard_Time</td><td>0.0105554191</td></tr>\n",
       "<tr><td>43</td><td>243</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/North_American_Eastern_Time_Zone</td><td>0.0098569008</td></tr>\n",
       "<tr><td>44</td><td>3</td><td>http://dbpedia.org/ontology/areaLand#count</td><td>0.0094568186</td></tr>\n",
       "<tr><td>45</td><td>4</td><td>http://dbpedia.org/ontology/areaLand#1</td><td>0.0094568186</td></tr>\n",
       "<tr><td>46</td><td>113</td><td>http://dbpedia.org/ontology/minimumElevation#count</td><td>0.0094335063</td></tr>\n",
       "<tr><td>47</td><td>118</td><td>http://dbpedia.org/ontology/populationAsOf#1@\"2006-12-31\"^^http://www.w3.org/2001/XMLSchema#date</td><td>0.0094283539</td></tr>\n",
       "<tr><td>48</td><td>261</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Village</td><td>0.0094071382</td></tr>\n",
       "<tr><td>49</td><td>114</td><td>http://dbpedia.org/ontology/minimumElevation#1</td><td>0.0093485650</td></tr>\n",
       "<tr><td>50</td><td>240</td><td>http://dbpedia.org/ontology/routeStart?inv#count</td><td>0.0092929410</td></tr>\n",
       "<tr><td>51</td><td>8</td><td>http://dbpedia.org/ontology/areaWater#count</td><td>0.0092498928</td></tr>\n",
       "<tr><td>52</td><td>258</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Census-designated_place</td><td>0.0089104877</td></tr>\n",
       "<tr><td>53</td><td>238</td><td>http://dbpedia.org/ontology/routeEnd?inv#count</td><td>0.0089041982</td></tr>\n",
       "<tr><td>54</td><td>111</td><td>http://dbpedia.org/ontology/maximumElevation#count</td><td>0.0087616361</td></tr>\n",
       "<tr><td>55</td><td>112</td><td>http://dbpedia.org/ontology/maximumElevation#1</td><td>0.0087616361</td></tr>\n",
       "<tr><td>56</td><td>287</td><td>http://dbpedia.org/ontology/utcOffset#1@\"-5\"</td><td>0.0082976925</td></tr>\n",
       "<tr><td>57</td><td>5</td><td>http://dbpedia.org/ontology/areaLand#2</td><td>0.0080500379</td></tr>\n",
       "<tr><td>58</td><td>133</td><td>http://dbpedia.org/ontology/populationDensity#count</td><td>0.0078484319</td></tr>\n",
       "<tr><td>59</td><td>6</td><td>http://dbpedia.org/ontology/areaTotal#count</td><td>0.0074483968</td></tr>\n",
       "<tr><td>60</td><td>304</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/List_of_twin_towns</td><td>0.0073464966</td></tr>\n",
       "<tr><td>61</td><td>294</td><td>http://dbpedia.org/ontology/utcOffset#2@\"-4\"</td><td>0.0067065538</td></tr>\n",
       "<tr><td>62</td><td>244</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/North_American_Central_Time_Zone</td><td>0.0066068794</td></tr>\n",
       "<tr><td>63</td><td>23</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Austria</td><td>0.0062625474</td></tr>\n",
       "<tr><td>64</td><td>116</td><td>http://dbpedia.org/ontology/nearestCity?inv#count</td><td>0.0059831857</td></tr>\n",
       "<tr><td>65</td><td>288</td><td>http://dbpedia.org/ontology/utcOffset#1@\"+2\"</td><td>0.0056993889</td></tr>\n",
       "<tr><td>66</td><td>297</td><td>http://dbpedia.org/ontology/utcOffset#2@\"+3\"</td><td>0.0053709065</td></tr>\n",
       "<tr><td>67</td><td>245</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/Eastern_European_Time</td><td>0.0053411230</td></tr>\n",
       "<tr><td>68</td><td>95</td><td>http://dbpedia.org/ontology/elevation#count</td><td>0.0052806194</td></tr>\n",
       "<tr><td>69</td><td>1</td><td>http://dbpedia.org/ontology/area#1</td><td>0.0051504748</td></tr>\n",
       "<tr><td>70</td><td>2</td><td>http://dbpedia.org/ontology/areaCode#count</td><td>0.0048780440</td></tr>\n",
       "<tr><td>71</td><td>96</td><td>http://dbpedia.org/ontology/elevation#1</td><td>0.0046697548</td></tr>\n",
       "<tr><td>72</td><td>279</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Municipalities_of_Brazil</td><td>0.0044755608</td></tr>\n",
       "<tr><td>73</td><td>15</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Germany</td><td>0.0043193264</td></tr>\n",
       "<tr><td>74</td><td>296</td><td>http://dbpedia.org/ontology/utcOffset#2@\"-6\"</td><td>0.0042388718</td></tr>\n",
       "<tr><td>75</td><td>259</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Town</td><td>0.0042273599</td></tr>\n",
       "<tr><td>76</td><td>135</td><td>http://dbpedia.org/ontology/postalCode#count</td><td>0.0040125311</td></tr>\n",
       "<tr><td>77</td><td>22</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Czech_Republic</td><td>0.0040068343</td></tr>\n",
       "<tr><td>78</td><td>115</td><td>http://dbpedia.org/ontology/motto#count</td><td>0.0038502787</td></tr>\n",
       "<tr><td>79</td><td>256</td><td>http://dbpedia.org/ontology/type#count</td><td>0.0037569686</td></tr>\n",
       "<tr><td>80</td><td>239</td><td>http://dbpedia.org/ontology/routeJunction?inv#count</td><td>0.0037468734</td></tr>\n",
       "<tr><td>81</td><td>254</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/Iran_Standard_Time</td><td>0.0036248488</td></tr>\n",
       "<tr><td>82</td><td>264</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Village_(United_States)</td><td>0.0036064034</td></tr>\n",
       "<tr><td>83</td><td>284</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Municipalities_of_Cuba</td><td>0.0035736682</td></tr>\n",
       "<tr><td>84</td><td>20</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Australia</td><td>0.0034511135</td></tr>\n",
       "<tr><td>85</td><td>18</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Poland</td><td>0.0033868709</td></tr>\n",
       "<tr><td>86</td><td>120</td><td>http://dbpedia.org/ontology/populationAsOf#1@\"2010-04-01\"^^http://www.w3.org/2001/XMLSchema#date</td><td>0.0033147498</td></tr>\n",
       "<tr><td>87</td><td>295</td><td>http://dbpedia.org/ontology/utcOffset#2@\"+1\"</td><td>0.0031843944</td></tr>\n",
       "<tr><td>88</td><td>242</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/Central_European_Time</td><td>0.0028993576</td></tr>\n",
       "<tr><td>89</td><td>285</td><td>http://dbpedia.org/ontology/type#1@OTHER</td><td>0.0026631391</td></tr>\n",
       "<tr><td>90</td><td>272</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/City_(California)</td><td>0.0025227055</td></tr>\n",
       "<tr><td>91</td><td>25</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Mexico</td><td>0.0024821676</td></tr>\n",
       "<tr><td>92</td><td>146</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Kyushu</td><td>0.0024320431</td></tr>\n",
       "<tr><td>93</td><td>107</td><td>http://dbpedia.org/ontology/leaderTitle#1@\"Municipal President\"@en</td><td>0.0022439285</td></tr>\n",
       "<tr><td>94</td><td>119</td><td>http://dbpedia.org/ontology/populationAsOf#1@\"2010-12-31\"^^http://www.w3.org/2001/XMLSchema#date</td><td>0.0021322853</td></tr>\n",
       "<tr><td>95</td><td>266</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Municipality</td><td>0.0019976217</td></tr>\n",
       "<tr><td>96</td><td>9</td><td>http://dbpedia.org/ontology/areaWater#1</td><td>0.0019811294</td></tr>\n",
       "<tr><td>97</td><td>148</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Chūbu_region</td><td>0.0019766421</td></tr>\n",
       "<tr><td>98</td><td>26</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Hungary</td><td>0.0019426290</td></tr>\n",
       "<tr><td>99</td><td>136</td><td>http://dbpedia.org/ontology/region#count</td><td>0.0019369811</td></tr>\n",
       "<tr><td>100</td><td>134</td><td>http://dbpedia.org/ontology/populationDensity#1</td><td>0.0019254055</td></tr>\n",
       "<tr><td>101</td><td>269</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Gmina</td><td>0.0019167945</td></tr>\n",
       "<tr><td>102</td><td>31</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Ukraine</td><td>0.0019097990</td></tr>\n",
       "<tr><td>103</td><td>28</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Sweden</td><td>0.0016357978</td></tr>\n",
       "<tr><td>104</td><td>302</td><td>http://www.w3.org/1999/02/22-rdf-syntax-ns#type#1@OTHER</td><td>0.0016116168</td></tr>\n",
       "<tr><td>105</td><td>16</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/United_Kingdom</td><td>0.0015910670</td></tr>\n",
       "<tr><td>106</td><td>163</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Kansai_region</td><td>0.0014567274</td></tr>\n",
       "<tr><td>107</td><td>122</td><td>http://dbpedia.org/ontology/populationAsOf#1@\"2009-06-30\"^^http://www.w3.org/2001/XMLSchema#date</td><td>0.0013583905</td></tr>\n",
       "<tr><td>108</td><td>12</td><td>http://dbpedia.org/ontology/country#count</td><td>0.0013441720</td></tr>\n",
       "<tr><td>109</td><td>293</td><td>http://dbpedia.org/ontology/utcOffset#1@OTHER</td><td>0.0012687612</td></tr>\n",
       "<tr><td>110</td><td>274</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Borough_(Pennsylvania)</td><td>0.0011648049</td></tr>\n",
       "<tr><td>111</td><td>139</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Pays_de_la_Loire</td><td>0.0010797894</td></tr>\n",
       "<tr><td>112</td><td>273</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/List_of_cities_in_Argentina</td><td>0.0010213124</td></tr>\n",
       "<tr><td>113</td><td>32</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/United_States_of_America</td><td>0.0010138343</td></tr>\n",
       "<tr><td>114</td><td>21</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Spain</td><td>0.0009926035</td></tr>\n",
       "<tr><td>115</td><td>308</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Ukraine</td><td>0.0009778459</td></tr>\n",
       "<tr><td>116</td><td>275</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Borough_(New_Jersey)</td><td>0.0009481932</td></tr>\n",
       "<tr><td>117</td><td>250</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/Mountain_Time_Zone</td><td>0.0008733979</td></tr>\n",
       "<tr><td>118</td><td>307</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Italy</td><td>0.0008481666</td></tr>\n",
       "<tr><td>119</td><td>137</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Nord-Pas-de-Calais</td><td>0.0007872309</td></tr>\n",
       "<tr><td>120</td><td>280</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Municipalities_of_Colombia</td><td>0.0007818404</td></tr>\n",
       "<tr><td>121</td><td>24</td><td>http://dbpedia.org/ontology/country#1@http://dbpedia.org/resource/Italy</td><td>0.0007681960</td></tr>\n",
       "<tr><td>122</td><td>255</td><td>http://dbpedia.org/ontology/timeZone#1@OTHER</td><td>0.0007568887</td></tr>\n",
       "<tr><td>123</td><td>251</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/Eastern_Standard_Time_Zone</td><td>0.0007361581</td></tr>\n",
       "<tr><td>124</td><td>263</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/New_England_town</td><td>0.0006881598</td></tr>\n",
       "<tr><td>125</td><td>270</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/List_of_towns_and_villages_in_Illinois</td><td>0.0006643873</td></tr>\n",
       "<tr><td>126</td><td>282</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/City_(Quebec)</td><td>0.0006639864</td></tr>\n",
       "<tr><td>127</td><td>314</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Africa</td><td>0.0006526736</td></tr>\n",
       "<tr><td>128</td><td>35</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Nord_(French_department)</td><td>0.0006488608</td></tr>\n",
       "<tr><td>129</td><td>140</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Rhône-Alpes</td><td>0.0006098502</td></tr>\n",
       "<tr><td>130</td><td>117</td><td>http://dbpedia.org/ontology/populationAsOf#count</td><td>0.0005952472</td></tr>\n",
       "<tr><td>131</td><td>257</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/City</td><td>0.0005771659</td></tr>\n",
       "<tr><td>132</td><td>141</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Aquitaine</td><td>0.0005468276</td></tr>\n",
       "<tr><td>133</td><td>40</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Maine-et-Loire</td><td>0.0005248101</td></tr>\n",
       "<tr><td>134</td><td>39</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Loire-Atlantique</td><td>0.0005043381</td></tr>\n",
       "<tr><td>135</td><td>142</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Lombardy</td><td>0.0005034930</td></tr>\n",
       "<tr><td>136</td><td>277</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Corregimientos_of_Panama</td><td>0.0004998276</td></tr>\n",
       "<tr><td>137</td><td>188</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Hokkaido</td><td>0.0004952871</td></tr>\n",
       "<tr><td>138</td><td>69</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Dordogne</td><td>0.0004947210</td></tr>\n",
       "<tr><td>139</td><td>271</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Administrative_divisions_of_New_York</td><td>0.0004838897</td></tr>\n",
       "<tr><td>140</td><td>41</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Haut-Rhin</td><td>0.0004605897</td></tr>\n",
       "<tr><td>141</td><td>316</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Asia</td><td>0.0004484091</td></tr>\n",
       "<tr><td>142</td><td>38</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Finistère</td><td>0.0004417426</td></tr>\n",
       "<tr><td>143</td><td>106</td><td>http://dbpedia.org/ontology/leaderTitle#1@\"Alcalde\"@en</td><td>0.0004371334</td></tr>\n",
       "<tr><td>144</td><td>227</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Düsseldorf</td><td>0.0004304264</td></tr>\n",
       "<tr><td>145</td><td>309</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/The_United_Kingdom</td><td>0.0004189674</td></tr>\n",
       "<tr><td>146</td><td>305</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Hungary</td><td>0.0004174443</td></tr>\n",
       "<tr><td>147</td><td>170</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Kassel</td><td>0.0004139833</td></tr>\n",
       "<tr><td>148</td><td>37</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Pas-de-Calais</td><td>0.0004096477</td></tr>\n",
       "<tr><td>149</td><td>248</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/Pacific_Time_Zone</td><td>0.0004072352</td></tr></table><h3>Features to drop</h3><table><tr><th>Position</th><th>Feat#</th><th>Feat</th><th>Utility</th></tr><tr><td>0</td><td>344</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Northern_Cyprus</td><td>0.0000040442</td></tr>\n",
       "<tr><td>1</td><td>126</td><td>http://dbpedia.org/ontology/populationAsOf#1@\"2008-12-31\"^^http://www.w3.org/2001/XMLSchema#date</td><td>0.0000080627</td></tr>\n",
       "<tr><td>2</td><td>371</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/History_of_Hamburg</td><td>0.0000092315</td></tr>\n",
       "<tr><td>3</td><td>207</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Chester_County,_Pennsylvania</td><td>0.0000104297</td></tr>\n",
       "<tr><td>4</td><td>222</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/McHenry_County,_Illinois</td><td>0.0000151230</td></tr>\n",
       "<tr><td>5</td><td>359</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Transport</td><td>0.0000181687</td></tr>\n",
       "<tr><td>6</td><td>125</td><td>http://dbpedia.org/ontology/populationAsOf#1@\"2007-12-31\"^^http://www.w3.org/2001/XMLSchema#date</td><td>0.0000185330</td></tr>\n",
       "<tr><td>7</td><td>206</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Madison_County,_Illinois</td><td>0.0000202727</td></tr>\n",
       "<tr><td>8</td><td>130</td><td>http://dbpedia.org/ontology/populationAsOf#1@\"2011-12-31\"^^http://www.w3.org/2001/XMLSchema#date</td><td>0.0000206101</td></tr>\n",
       "<tr><td>9</td><td>199</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Piedmont</td><td>0.0000232673</td></tr>\n",
       "<tr><td>10</td><td>214</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Marche</td><td>0.0000247764</td></tr>\n",
       "<tr><td>11</td><td>231</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Lackawanna_County,_Pennsylvania</td><td>0.0000255142</td></tr>\n",
       "<tr><td>12</td><td>219</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Trenčín_Region</td><td>0.0000261496</td></tr>\n",
       "<tr><td>13</td><td>83</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Loire_(département)</td><td>0.0000262310</td></tr>\n",
       "<tr><td>14</td><td>131</td><td>http://dbpedia.org/ontology/populationAsOf#1@\"2014-12-31\"^^http://www.w3.org/2001/XMLSchema#date</td><td>0.0000294894</td></tr>\n",
       "<tr><td>15</td><td>196</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Will_County,_Illinois</td><td>0.0000323192</td></tr>\n",
       "<tr><td>16</td><td>176</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Moravian-Silesian_Region</td><td>0.0000324838</td></tr>\n",
       "<tr><td>17</td><td>224</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Capital_District</td><td>0.0000332389</td></tr>\n",
       "<tr><td>18</td><td>252</td><td>http://dbpedia.org/ontology/timeZone#1@http://dbpedia.org/resource/Central_Time_Zone_(North_America)</td><td>0.0000341756</td></tr>\n",
       "<tr><td>19</td><td>50</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Seine-Maritime</td><td>0.0000341793</td></tr>\n",
       "<tr><td>20</td><td>235</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Hudson_Valley</td><td>0.0000346144</td></tr>\n",
       "<tr><td>21</td><td>232</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Darmstadt_(region)</td><td>0.0000347087</td></tr>\n",
       "<tr><td>22</td><td>208</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/St._Clair_County,_Illinois</td><td>0.0000352008</td></tr>\n",
       "<tr><td>23</td><td>362</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Banstead</td><td>0.0000365581</td></tr>\n",
       "<tr><td>24</td><td>361</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Malta</td><td>0.0000365581</td></tr>\n",
       "<tr><td>25</td><td>364</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Marion_County</td><td>0.0000369277</td></tr>\n",
       "<tr><td>26</td><td>366</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Oceania</td><td>0.0000369800</td></tr>\n",
       "<tr><td>27</td><td>363</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Montenegro</td><td>0.0000369800</td></tr>\n",
       "<tr><td>28</td><td>332</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Slovenia</td><td>0.0000370154</td></tr>\n",
       "<tr><td>29</td><td>194</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/South_East_England</td><td>0.0000371703</td></tr>\n",
       "<tr><td>30</td><td>368</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/List_of_people_from_the_Pittsburgh_metropolitan_area</td><td>0.0000372525</td></tr>\n",
       "<tr><td>31</td><td>36</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Yvelines</td><td>0.0000414433</td></tr>\n",
       "<tr><td>32</td><td>77</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Val-d'Oise</td><td>0.0000417613</td></tr>\n",
       "<tr><td>33</td><td>76</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Drôme</td><td>0.0000417613</td></tr>\n",
       "<tr><td>34</td><td>221</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Jefferson_Parish,_Louisiana</td><td>0.0000482626</td></tr>\n",
       "<tr><td>35</td><td>220</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/East_Midlands</td><td>0.0000482626</td></tr>\n",
       "<tr><td>36</td><td>223</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Sangamon_County,_Illinois</td><td>0.0000496529</td></tr>\n",
       "<tr><td>37</td><td>53</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Nord_(département)</td><td>0.0000502507</td></tr>\n",
       "<tr><td>38</td><td>72</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Ain</td><td>0.0000510437</td></tr>\n",
       "<tr><td>39</td><td>84</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Côte-d'Or</td><td>0.0000514225</td></tr>\n",
       "<tr><td>40</td><td>187</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Champagne-Ardenne</td><td>0.0000514636</td></tr>\n",
       "<tr><td>41</td><td>218</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Trnava_Region</td><td>0.0000517374</td></tr>\n",
       "<tr><td>42</td><td>202</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Olomouc_Region</td><td>0.0000523276</td></tr>\n",
       "<tr><td>43</td><td>67</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Alpes-Maritimes</td><td>0.0000526531</td></tr>\n",
       "<tr><td>44</td><td>92</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Haute-Vienne</td><td>0.0000526967</td></tr>\n",
       "<tr><td>45</td><td>90</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Marne_(département)</td><td>0.0000526967</td></tr>\n",
       "<tr><td>46</td><td>228</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Žilina_Region</td><td>0.0000545172</td></tr>\n",
       "<tr><td>47</td><td>320</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Lithuania</td><td>0.0000555288</td></tr>\n",
       "<tr><td>48</td><td>91</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Allier</td><td>0.0000557335</td></tr>\n",
       "<tr><td>49</td><td>230</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Madeira_Island</td><td>0.0000561838</td></tr>\n",
       "<tr><td>50</td><td>197</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/North_West_England</td><td>0.0000585648</td></tr>\n",
       "<tr><td>51</td><td>145</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Campania</td><td>0.0000592991</td></tr>\n",
       "<tr><td>52</td><td>205</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Apulia</td><td>0.0000604698</td></tr>\n",
       "<tr><td>53</td><td>129</td><td>http://dbpedia.org/ontology/populationAsOf#1@\"2009-12-31\"^^http://www.w3.org/2001/XMLSchema#date</td><td>0.0000608310</td></tr>\n",
       "<tr><td>54</td><td>281</td><td>http://dbpedia.org/ontology/type#1@http://dbpedia.org/resource/Municipalities_of_Mexico</td><td>0.0000621425</td></tr>\n",
       "<tr><td>55</td><td>87</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Moselle_(département)</td><td>0.0000627847</td></tr>\n",
       "<tr><td>56</td><td>167</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Sicily</td><td>0.0000630973</td></tr>\n",
       "<tr><td>57</td><td>318</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Estonia</td><td>0.0000636066</td></tr>\n",
       "<tr><td>58</td><td>73</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Var_(département)</td><td>0.0000646552</td></tr>\n",
       "<tr><td>59</td><td>323</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/The_Republic_of_Ireland</td><td>0.0000689172</td></tr>\n",
       "<tr><td>60</td><td>212</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Île-de-France</td><td>0.0000693302</td></tr>\n",
       "<tr><td>61</td><td>192</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Karlovy_Vary_Region</td><td>0.0000695792</td></tr>\n",
       "<tr><td>62</td><td>61</td><td>http://dbpedia.org/ontology/department#1@http://dbpedia.org/resource/Vaucluse</td><td>0.0000696101</td></tr>\n",
       "<tr><td>63</td><td>233</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Abruzzo</td><td>0.0000698774</td></tr>\n",
       "<tr><td>64</td><td>322</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Latvia</td><td>0.0000700137</td></tr>\n",
       "<tr><td>65</td><td>355</td><td>http://www.w3.org/2000/01/rdf-schema#seeAlso#3@http://dbpedia.org/resource/Media</td><td>0.0000705965</td></tr>\n",
       "<tr><td>66</td><td>234</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Latium</td><td>0.0000707228</td></tr>\n",
       "<tr><td>67</td><td>147</td><td>http://dbpedia.org/ontology/region#1@http://dbpedia.org/resource/Veneto</td><td>0.0000717788</td></tr></table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 28\n",
    "import numpy as np\n",
    "import pickle\n",
    "import random\n",
    "from collections import OrderedDict\n",
    "\n",
    "with open(\"ch6_cell27_splits.pk\", \"rb\") as pkl:\n",
    "    segments_at = pickle.load(pkl)\n",
    "\n",
    "data = list()\n",
    "header = None\n",
    "PARAM_RANDOM_FEATS = 10\n",
    "rand = random.Random(42)\n",
    "with open(\"ch6_cell26_dev_feat1_dampen.tsv\") as feats:\n",
    "    header = next(feats)\n",
    "    header = header.strip().split(\"\\t\")\n",
    "    header.pop(0) # name\n",
    "    header.pop() # population\n",
    "    for idx in range(PARAM_RANDOM_FEATS):\n",
    "        header.append('random_' + str(idx))\n",
    "    for line in feats:\n",
    "        fields = line.strip().split(\"\\t\")\n",
    "        logpop = float(fields[-1])\n",
    "        name = fields[0]\n",
    "        feats = list(map(float,fields[1:-1]))\n",
    "        for _ in range(PARAM_RANDOM_FEATS):\n",
    "            feats.append(rand.random())\n",
    "        row = (feats, logpop, name)\n",
    "        data.append(row)\n",
    "\n",
    "xdata = np.array(list(map(lambda t:t[0], data)))\n",
    "ydata = np.array(list(map(lambda t:t[1], data)))\n",
    "\n",
    "def cell28_adjudicate(data, segments):\n",
    "    result = list()\n",
    "    for val in data:\n",
    "        idx = None\n",
    "        if val < segments[0]['min']:\n",
    "            idx = 0\n",
    "        elif val > segments[-1]['max']:\n",
    "            idx = len(segments) - 1\n",
    "        else:\n",
    "            for idx2, segment in enumerate(segments):\n",
    "#                if (segment['min'] == val and segment['max'] == val) or \\\n",
    "#                    (segment['min'] <= val and \\\n",
    "#                    (idx2 == len(segments)-1 or val < segments[idx2+1]['min'])):\n",
    "                if segment['min'] <= val and val <= segment['max']:\n",
    "                    idx = idx2\n",
    "                    break\n",
    "        result.append(idx)\n",
    "    return np.array(result)\n",
    "\n",
    "ydata = cell28_adjudicate(ydata, segments_at[2])\n",
    "\n",
    "# discretize features into up to 4 bins\n",
    "for feat in range(xdata.shape[1]):\n",
    "    feat_segments = cell27_segment(sorted(xdata[:,feat]), 3)\n",
    "    xdata[:,feat] = cell28_adjudicate(xdata[:,feat], feat_segments[2])\n",
    "    \n",
    "# compute confusion tables for each feature\n",
    "confusion_tables = list()\n",
    "for feat in range(xdata.shape[1]):\n",
    "    table = OrderedDict()\n",
    "    for row in range(xdata.shape[0]):\n",
    "        feat_val   = int(xdata[row, feat])\n",
    "        target_val = int(ydata[row])\n",
    "        if feat_val not in table:\n",
    "            table[feat_val] = OrderedDict()\n",
    "        table[feat_val][target_val] = table[feat_val].get(target_val, 0) + 1\n",
    "    confusion_tables.append(table)\n",
    "\n",
    "# compute feature utility using mutual information\n",
    "PARAM_USE_CHI = False\n",
    "feature_utility = list()\n",
    "if PARAM_USE_CHI:\n",
    "    # pip install scipy\n",
    "    from scipy.stats.mstats import chisquare\n",
    "for feat in range(xdata.shape[1]):\n",
    "    if feat % int(xdata.shape[1] / 10) == 0:\n",
    "        print(\"{:>3} out of {:>3}\".format(feat, xdata.shape[1]))\n",
    "    table = confusion_tables[feat]\n",
    "    feat_vals = set()\n",
    "    for row in table.values():\n",
    "        feat_vals.update(row.keys())\n",
    "    cols = { val: sum(map(lambda x:x.get(val,0), table.values())) for val in feat_vals }\n",
    "    full_table = sum(cols.values())\n",
    "    \n",
    "    best_utility = None\n",
    "    for feat_val in table.keys():\n",
    "        for target_val in table[feat_val].keys():\n",
    "            # binarize\n",
    "            n11 = table[feat_val][target_val]\n",
    "            if n11 < 5:\n",
    "                continue\n",
    "            n10 = sum(table[feat_val].values()) - n11\n",
    "            n01 = cols.get(target_val) - n11\n",
    "            n00 = full_table - n11 - n10 - n01\n",
    "            if n10 == 0 or n01 == 0 or n00 == 0:\n",
    "                continue\n",
    "            if PARAM_USE_CHI:\n",
    "                utility, _ = chisquare([n11, n10, n01, n00])\n",
    "            else:\n",
    "                n1_ = n11 + n10\n",
    "                n0_ = n01 + n00\n",
    "                n_1 = n11 + n01\n",
    "                n_0 = n10 + n00\n",
    "                n = float(full_table)\n",
    "                utility = n11/n * math.log(n*n11/(n1_*n_1),2) + \\\n",
    "                   n01 / n * math.log(n*n01/(n0_*n_1), 2) + \\\n",
    "                   n10 / n * math.log(n*n10/(n1_*n_0), 2) + \\\n",
    "                   n00 / n * math.log(n*n00/(n0_*n_0), 2)\n",
    "            if best_utility is None or best_utility < utility:\n",
    "                best_utility = utility\n",
    "    if best_utility is not None:\n",
    "        feature_utility.append( (feat, best_utility) )\n",
    "    #else:\n",
    "    #    print(\"No utility for feat \", feat)\n",
    "    \n",
    "feature_utility = sorted(feature_utility, key=lambda x:x[1], reverse=True)\n",
    "\n",
    "# determine features to drop based on whether there are PARAM_BEST_RANDOM on top of them\n",
    "PARAM_BEST_RANDOM = 3\n",
    "cut_point = None\n",
    "random_so_far = 0\n",
    "for idx, row in enumerate(feature_utility):\n",
    "    if header[row[0]].startswith(\"random_\"):\n",
    "        random_so_far += 1\n",
    "        if random_so_far == PARAM_BEST_RANDOM:\n",
    "            cut_point = idx\n",
    "            break\n",
    "            \n",
    "to_drop = list()\n",
    "for row in feature_utility[cut_point:]:\n",
    "    if not header[row[0]].startswith(\"random_\"):\n",
    "        to_drop.append(row)\n",
    "        \n",
    "print(\"Dropped\", len(to_drop), \"features\")\n",
    "\n",
    "to_keep = set()\n",
    "with open(\"ch6_cell28_features_dropped.txt\", \"w\") as dropped:\n",
    "    drop_set = set()\n",
    "    for row in to_drop:\n",
    "        dropped.write(header[row[0]] + \"\\n\")\n",
    "        drop_set.add(row[0])\n",
    "    for idx in range(len(header) - PARAM_RANDOM_FEATS):\n",
    "        if idx not in drop_set:\n",
    "            to_keep.add(idx + 1) # accomodate 'name' column\n",
    "        to_keep.add(0)\n",
    "        to_keep.add(len(header) - PARAM_RANDOM_FEATS + 1) # logpop\n",
    "\n",
    "print(\"New feature vector size\", len(to_keep), \"features, including name and target\")\n",
    "        \n",
    "table1 = (\"<table><tr><th>Position</th><th>Feat#</th><th>Feat</th><th>Utility</th></tr>\" +\n",
    "            \"\\n\".join(list(map(lambda r: \n",
    "                               \"<tr><td>{}</td><td>{}</td><td>{}</td><td>{:5.10f}</td></tr>\".format(\n",
    "                        r[0], r[1][0],\n",
    "                        header[r[1][0]].replace('<','').replace('>',''),\n",
    "                        r[1][1]), \n",
    "                               enumerate(feature_utility[:150])))) +\"</table>\")\n",
    "table2 = (\"<table><tr><th>Position</th><th>Feat#</th><th>Feat</th><th>Utility</th></tr>\" +\n",
    "            \"\\n\".join(list(map(lambda r: \n",
    "                               \"<tr><td>{}</td><td>{}</td><td>{}</td><td>{:5.10f}</td></tr>\".format(\n",
    "                        r[0], r[1][0],\n",
    "                        header[r[1][0]].replace('<','').replace('>',''),\n",
    "                        r[1][1]), \n",
    "                               enumerate(reversed(to_drop))))) +\"</table>\")\n",
    "\n",
    "with open(\"ch6_cell28_dev_feat1_filtered.tsv\", \"w\") as filtered:\n",
    "    with open(\"ch6_cell26_dev_feat1_dampen.tsv\") as feats:\n",
    "        header = next(feats)\n",
    "        header = header.strip().split(\"\\t\")\n",
    "        new_header = list()\n",
    "        for idx, header_col in enumerate(header):\n",
    "            if idx in to_keep:\n",
    "                new_header.append(header_col)\n",
    "        filtered.write(\"\\t\".join(new_header) + \"\\n\")\n",
    "        for line in feats:\n",
    "            fields = line.strip().split(\"\\t\")\n",
    "            new_fields = list()\n",
    "            for idx, value in enumerate(fields):\n",
    "                if idx in to_keep:\n",
    "                    new_fields.append(value)\n",
    "            filtered.write(\"\\t\".join(new_fields) + \"\\n\")\n",
    "\n",
    "from IPython.display import HTML, display\n",
    "display(HTML(\"<h3>Top 150 features by MI</h3>\" + table1 + \n",
    "             \"<h3>Features to drop</h3>\"       + table2))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The list of most informative features look very good, particularly number of timezones and actual timezones indicates the geographic location is important. While the wikipedia page contains the GPS coordinates of each city, that information is available in the dbpedia files I chose, but it is available in the geonames file, so it would be useful to retrieve it. I'll look into it in the GIS case study in Chapter 10.\n",
    "\n",
    "The counts of incoming links seem also very useful, together with the areaTotal. The features dropped seem very specific and most probably fire very little.\n",
    "\n",
    "I will not turn to another technique for analyzing the data, that is to use decision trees using the splits. Cell #29"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/svg+xml": [
       "<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"no\"?>\n",
       "<!DOCTYPE svg PUBLIC \"-//W3C//DTD SVG 1.1//EN\"\n",
       " \"http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd\">\n",
       "<!-- Generated by graphviz version 2.43.0 (0)\n",
       " -->\n",
       "<!-- Title: Tree Pages: 1 -->\n",
       "<svg width=\"3293pt\" height=\"489pt\"\n",
       " viewBox=\"0.00 0.00 3293.00 489.00\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\">\n",
       "<g id=\"graph0\" class=\"graph\" transform=\"scale(1 1) rotate(0) translate(4 485)\">\n",
       "<title>Tree</title>\n",
       "<polygon fill=\"white\" stroke=\"transparent\" points=\"-4,4 -4,-485 3289,-485 3289,4 -4,4\"/>\n",
       "<!-- 0 -->\n",
       "<g id=\"node1\" class=\"node\">\n",
       "<title>0</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1830,-481 1598,-481 1598,-428 1830,-428 1830,-481\"/>\n",
       "<text text-anchor=\"middle\" x=\"1714\" y=\"-465.8\" font-family=\"Times,serif\" font-size=\"14.00\">birthPlace?inv#count &lt;= 0.739</text>\n",
       "<text text-anchor=\"middle\" x=\"1714\" y=\"-450.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 44959</text>\n",
       "<text text-anchor=\"middle\" x=\"1714\" y=\"-435.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [11240, 11172, 11318, 11229]</text>\n",
       "</g>\n",
       "<!-- 1 -->\n",
       "<g id=\"node2\" class=\"node\">\n",
       "<title>1</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1385,-392 1167,-392 1167,-339 1385,-339 1385,-392\"/>\n",
       "<text text-anchor=\"middle\" x=\"1276\" y=\"-376.8\" font-family=\"Times,serif\" font-size=\"14.00\">country#1@THER &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"1276\" y=\"-361.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 37696</text>\n",
       "<text text-anchor=\"middle\" x=\"1276\" y=\"-346.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [11042, 10545, 9236, 6873]</text>\n",
       "</g>\n",
       "<!-- 0&#45;&gt;1 -->\n",
       "<g id=\"edge1\" class=\"edge\">\n",
       "<title>0&#45;&gt;1</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1597.77,-430.41C1535.23,-417.99 1458.21,-402.69 1395.21,-390.18\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1395.72,-386.71 1385.23,-388.2 1394.36,-393.58 1395.72,-386.71\"/>\n",
       "<text text-anchor=\"middle\" x=\"1399.12\" y=\"-405.28\" font-family=\"Times,serif\" font-size=\"14.00\">True</text>\n",
       "</g>\n",
       "<!-- 22042 -->\n",
       "<g id=\"node33\" class=\"node\">\n",
       "<title>22042</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2233.5,-392 2042.5,-392 2042.5,-339 2233.5,-339 2233.5,-392\"/>\n",
       "<text text-anchor=\"middle\" x=\"2138\" y=\"-376.8\" font-family=\"Times,serif\" font-size=\"14.00\">areaTotal#1 &lt;= 41230000.0</text>\n",
       "<text text-anchor=\"middle\" x=\"2138\" y=\"-361.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 7263</text>\n",
       "<text text-anchor=\"middle\" x=\"2138\" y=\"-346.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [198, 627, 2082, 4356]</text>\n",
       "</g>\n",
       "<!-- 0&#45;&gt;22042 -->\n",
       "<g id=\"edge32\" class=\"edge\">\n",
       "<title>0&#45;&gt;22042</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1830.14,-429.67C1893.31,-416.71 1970.87,-400.79 2032.33,-388.18\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2033.35,-391.55 2042.44,-386.11 2031.94,-384.69 2033.35,-391.55\"/>\n",
       "<text text-anchor=\"middle\" x=\"2028.68\" y=\"-403.28\" font-family=\"Times,serif\" font-size=\"14.00\">False</text>\n",
       "</g>\n",
       "<!-- 2 -->\n",
       "<g id=\"node3\" class=\"node\">\n",
       "<title>2</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"723.5,-303 518.5,-303 518.5,-250 723.5,-250 723.5,-303\"/>\n",
       "<text text-anchor=\"middle\" x=\"621\" y=\"-287.8\" font-family=\"Times,serif\" font-size=\"14.00\">country#1@India &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"621\" y=\"-272.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 28716</text>\n",
       "<text text-anchor=\"middle\" x=\"621\" y=\"-257.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [9680, 9017, 7032, 2987]</text>\n",
       "</g>\n",
       "<!-- 1&#45;&gt;2 -->\n",
       "<g id=\"edge2\" class=\"edge\">\n",
       "<title>1&#45;&gt;2</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1166.74,-349.99C1046.72,-334.05 854.95,-308.57 733.62,-292.46\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"733.93,-288.97 723.55,-291.12 733,-295.91 733.93,-288.97\"/>\n",
       "</g>\n",
       "<!-- 16465 -->\n",
       "<g id=\"node18\" class=\"node\">\n",
       "<title>16465</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1378.5,-303 1173.5,-303 1173.5,-250 1378.5,-250 1378.5,-303\"/>\n",
       "<text text-anchor=\"middle\" x=\"1276\" y=\"-287.8\" font-family=\"Times,serif\" font-size=\"14.00\">utcOffset#count &lt;= 0.389</text>\n",
       "<text text-anchor=\"middle\" x=\"1276\" y=\"-272.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 8980</text>\n",
       "<text text-anchor=\"middle\" x=\"1276\" y=\"-257.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [1362, 1528, 2204, 3886]</text>\n",
       "</g>\n",
       "<!-- 1&#45;&gt;16465 -->\n",
       "<g id=\"edge17\" class=\"edge\">\n",
       "<title>1&#45;&gt;16465</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1276,-338.87C1276,-330.89 1276,-321.93 1276,-313.38\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1279.5,-313.25 1276,-303.25 1272.5,-313.25 1279.5,-313.25\"/>\n",
       "</g>\n",
       "<!-- 3 -->\n",
       "<g id=\"node4\" class=\"node\">\n",
       "<title>3</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"421.5,-214 216.5,-214 216.5,-161 421.5,-161 421.5,-214\"/>\n",
       "<text text-anchor=\"middle\" x=\"319\" y=\"-198.8\" font-family=\"Times,serif\" font-size=\"14.00\">areaTotal#1 &lt;= 21493451.0</text>\n",
       "<text text-anchor=\"middle\" x=\"319\" y=\"-183.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 26957</text>\n",
       "<text text-anchor=\"middle\" x=\"319\" y=\"-168.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [9672, 8962, 6419, 1904]</text>\n",
       "</g>\n",
       "<!-- 2&#45;&gt;3 -->\n",
       "<g id=\"edge3\" class=\"edge\">\n",
       "<title>2&#45;&gt;3</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M532.43,-249.98C496.32,-239.58 454.43,-227.52 417.38,-216.84\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"418.15,-213.42 407.57,-214.01 416.21,-220.15 418.15,-213.42\"/>\n",
       "</g>\n",
       "<!-- 15336 -->\n",
       "<g id=\"node11\" class=\"node\">\n",
       "<title>15336</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"703,-214 539,-214 539,-161 703,-161 703,-214\"/>\n",
       "<text text-anchor=\"middle\" x=\"621\" y=\"-198.8\" font-family=\"Times,serif\" font-size=\"14.00\">rel#count &lt;= 1.097</text>\n",
       "<text text-anchor=\"middle\" x=\"621\" y=\"-183.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 1759</text>\n",
       "<text text-anchor=\"middle\" x=\"621\" y=\"-168.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [8, 55, 613, 1083]</text>\n",
       "</g>\n",
       "<!-- 2&#45;&gt;15336 -->\n",
       "<g id=\"edge10\" class=\"edge\">\n",
       "<title>2&#45;&gt;15336</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M621,-249.87C621,-241.89 621,-232.93 621,-224.38\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"624.5,-224.25 621,-214.25 617.5,-224.25 624.5,-224.25\"/>\n",
       "</g>\n",
       "<!-- 4 -->\n",
       "<g id=\"node5\" class=\"node\">\n",
       "<title>4</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"198,-125 0,-125 0,-72 198,-72 198,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"99\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">elevation#1 &lt;= 111.931</text>\n",
       "<text text-anchor=\"middle\" x=\"99\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 19403</text>\n",
       "<text text-anchor=\"middle\" x=\"99\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [7991, 6647, 4035, 730]</text>\n",
       "</g>\n",
       "<!-- 3&#45;&gt;4 -->\n",
       "<g id=\"edge4\" class=\"edge\">\n",
       "<title>3&#45;&gt;4</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M254.48,-160.98C228.96,-150.89 199.48,-139.23 173.09,-128.8\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"174.11,-125.44 163.52,-125.01 171.53,-131.95 174.11,-125.44\"/>\n",
       "</g>\n",
       "<!-- 10765 -->\n",
       "<g id=\"node8\" class=\"node\">\n",
       "<title>10765</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"421.5,-125 216.5,-125 216.5,-72 421.5,-72 421.5,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"319\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">country#1@Brazil &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"319\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 7554</text>\n",
       "<text text-anchor=\"middle\" x=\"319\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [1681, 2315, 2384, 1174]</text>\n",
       "</g>\n",
       "<!-- 3&#45;&gt;10765 -->\n",
       "<g id=\"edge7\" class=\"edge\">\n",
       "<title>3&#45;&gt;10765</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M319,-160.87C319,-152.89 319,-143.93 319,-135.38\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"322.5,-135.25 319,-125.25 315.5,-135.25 322.5,-135.25\"/>\n",
       "</g>\n",
       "<!-- 5 -->\n",
       "<g id=\"node6\" class=\"node\">\n",
       "<title>5</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"90,-36 36,-36 36,0 90,0 90,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"63\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 4&#45;&gt;5 -->\n",
       "<g id=\"edge5\" class=\"edge\">\n",
       "<title>4&#45;&gt;5</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M87.21,-71.8C83.34,-63.35 79.03,-53.95 75.13,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"78.3,-43.96 70.95,-36.33 71.93,-46.88 78.3,-43.96\"/>\n",
       "</g>\n",
       "<!-- 6970 -->\n",
       "<g id=\"node7\" class=\"node\">\n",
       "<title>6970</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"162,-36 108,-36 108,0 162,0 162,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"135\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 4&#45;&gt;6970 -->\n",
       "<g id=\"edge6\" class=\"edge\">\n",
       "<title>4&#45;&gt;6970</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M110.79,-71.8C114.66,-63.35 118.97,-53.95 122.87,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"126.07,-46.88 127.05,-36.33 119.7,-43.96 126.07,-46.88\"/>\n",
       "</g>\n",
       "<!-- 10766 -->\n",
       "<g id=\"node9\" class=\"node\">\n",
       "<title>10766</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"326,-36 272,-36 272,0 326,0 326,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"299\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 10765&#45;&gt;10766 -->\n",
       "<g id=\"edge8\" class=\"edge\">\n",
       "<title>10765&#45;&gt;10766</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M312.45,-71.8C310.37,-63.62 308.06,-54.55 305.95,-46.27\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"309.28,-45.16 303.42,-36.33 302.49,-46.88 309.28,-45.16\"/>\n",
       "</g>\n",
       "<!-- 15255 -->\n",
       "<g id=\"node10\" class=\"node\">\n",
       "<title>15255</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"398,-36 344,-36 344,0 398,0 398,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"371\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 10765&#45;&gt;15255 -->\n",
       "<g id=\"edge9\" class=\"edge\">\n",
       "<title>10765&#45;&gt;15255</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M336.03,-71.8C341.74,-63.17 348.12,-53.54 353.84,-44.9\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"356.92,-46.6 359.52,-36.33 351.08,-42.73 356.92,-46.6\"/>\n",
       "</g>\n",
       "<!-- 15337 -->\n",
       "<g id=\"node12\" class=\"node\">\n",
       "<title>15337</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"617,-125 439,-125 439,-72 617,-72 617,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"528\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">postalCode#count &lt;= 0.151</text>\n",
       "<text text-anchor=\"middle\" x=\"528\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 916</text>\n",
       "<text text-anchor=\"middle\" x=\"528\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [6, 41, 462, 407]</text>\n",
       "</g>\n",
       "<!-- 15336&#45;&gt;15337 -->\n",
       "<g id=\"edge11\" class=\"edge\">\n",
       "<title>15336&#45;&gt;15337</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M593.6,-160.87C584.04,-151.92 573.17,-141.75 563.07,-132.3\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"565.22,-129.53 555.53,-125.25 560.44,-134.64 565.22,-129.53\"/>\n",
       "</g>\n",
       "<!-- 16104 -->\n",
       "<g id=\"node15\" class=\"node\">\n",
       "<title>16104</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"793,-125 635,-125 635,-72 793,-72 793,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"714\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">rel#count &lt;= 1.19</text>\n",
       "<text text-anchor=\"middle\" x=\"714\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 843</text>\n",
       "<text text-anchor=\"middle\" x=\"714\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [2, 14, 151, 676]</text>\n",
       "</g>\n",
       "<!-- 15336&#45;&gt;16104 -->\n",
       "<g id=\"edge14\" class=\"edge\">\n",
       "<title>15336&#45;&gt;16104</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M648.4,-160.87C657.96,-151.92 668.83,-141.75 678.93,-132.3\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"681.56,-134.64 686.47,-125.25 676.78,-129.53 681.56,-134.64\"/>\n",
       "</g>\n",
       "<!-- 15338 -->\n",
       "<g id=\"node13\" class=\"node\">\n",
       "<title>15338</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"502,-36 448,-36 448,0 502,0 502,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"475\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 15337&#45;&gt;15338 -->\n",
       "<g id=\"edge12\" class=\"edge\">\n",
       "<title>15337&#45;&gt;15338</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M510.65,-71.8C504.76,-63.08 498.19,-53.34 492.3,-44.63\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"495.2,-42.66 486.7,-36.33 489.4,-46.58 495.2,-42.66\"/>\n",
       "</g>\n",
       "<!-- 15721 -->\n",
       "<g id=\"node14\" class=\"node\">\n",
       "<title>15721</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"574,-36 520,-36 520,0 574,0 574,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"547\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 15337&#45;&gt;15721 -->\n",
       "<g id=\"edge13\" class=\"edge\">\n",
       "<title>15337&#45;&gt;15721</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M534.22,-71.8C536.2,-63.62 538.39,-54.55 540.4,-46.27\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"543.85,-46.87 542.81,-36.33 537.05,-45.23 543.85,-46.87\"/>\n",
       "</g>\n",
       "<!-- 16105 -->\n",
       "<g id=\"node16\" class=\"node\">\n",
       "<title>16105</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"705,-36 651,-36 651,0 705,0 705,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"678\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 16104&#45;&gt;16105 -->\n",
       "<g id=\"edge15\" class=\"edge\">\n",
       "<title>16104&#45;&gt;16105</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M702.21,-71.8C698.34,-63.35 694.03,-53.95 690.13,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"693.3,-43.96 685.95,-36.33 686.93,-46.88 693.3,-43.96\"/>\n",
       "</g>\n",
       "<!-- 16374 -->\n",
       "<g id=\"node17\" class=\"node\">\n",
       "<title>16374</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"777,-36 723,-36 723,0 777,0 777,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"750\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 16104&#45;&gt;16374 -->\n",
       "<g id=\"edge16\" class=\"edge\">\n",
       "<title>16104&#45;&gt;16374</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M725.79,-71.8C729.66,-63.35 733.97,-53.95 737.87,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"741.07,-46.88 742.05,-36.33 734.7,-43.96 741.07,-46.88\"/>\n",
       "</g>\n",
       "<!-- 16466 -->\n",
       "<g id=\"node19\" class=\"node\">\n",
       "<title>16466</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1267.5,-214 1028.5,-214 1028.5,-161 1267.5,-161 1267.5,-214\"/>\n",
       "<text text-anchor=\"middle\" x=\"1148\" y=\"-198.8\" font-family=\"Times,serif\" font-size=\"14.00\">22&#45;rdf&#45;syntax&#45;ns#type#1@City &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"1148\" y=\"-183.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 6352</text>\n",
       "<text text-anchor=\"middle\" x=\"1148\" y=\"-168.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [692, 809, 1474, 3377]</text>\n",
       "</g>\n",
       "<!-- 16465&#45;&gt;16466 -->\n",
       "<g id=\"edge18\" class=\"edge\">\n",
       "<title>16465&#45;&gt;16466</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1238.29,-249.87C1224.48,-240.48 1208.69,-229.75 1194.23,-219.92\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1196.12,-216.98 1185.88,-214.25 1192.19,-222.77 1196.12,-216.98\"/>\n",
       "</g>\n",
       "<!-- 20285 -->\n",
       "<g id=\"node26\" class=\"node\">\n",
       "<title>20285</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1524.5,-214 1285.5,-214 1285.5,-161 1524.5,-161 1524.5,-214\"/>\n",
       "<text text-anchor=\"middle\" x=\"1405\" y=\"-198.8\" font-family=\"Times,serif\" font-size=\"14.00\">22&#45;rdf&#45;syntax&#45;ns#type#1@City &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"1405\" y=\"-183.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 2628</text>\n",
       "<text text-anchor=\"middle\" x=\"1405\" y=\"-168.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [670, 719, 730, 509]</text>\n",
       "</g>\n",
       "<!-- 16465&#45;&gt;20285 -->\n",
       "<g id=\"edge25\" class=\"edge\">\n",
       "<title>16465&#45;&gt;20285</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1314.01,-249.87C1327.92,-240.48 1343.83,-229.75 1358.41,-219.92\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1360.49,-222.74 1366.82,-214.25 1356.57,-216.94 1360.49,-222.74\"/>\n",
       "</g>\n",
       "<!-- 16467 -->\n",
       "<g id=\"node20\" class=\"node\">\n",
       "<title>16467</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1002.5,-125 811.5,-125 811.5,-72 1002.5,-72 1002.5,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"907\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">areaTotal#1 &lt;= 15925000.0</text>\n",
       "<text text-anchor=\"middle\" x=\"907\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 5251</text>\n",
       "<text text-anchor=\"middle\" x=\"907\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [682, 780, 1357, 2432]</text>\n",
       "</g>\n",
       "<!-- 16466&#45;&gt;16467 -->\n",
       "<g id=\"edge19\" class=\"edge\">\n",
       "<title>16466&#45;&gt;16467</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1077.32,-160.98C1049.12,-150.8 1016.5,-139.03 987.4,-128.52\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"988.27,-125.12 977.68,-125.01 985.9,-131.7 988.27,-125.12\"/>\n",
       "</g>\n",
       "<!-- 19970 -->\n",
       "<g id=\"node23\" class=\"node\">\n",
       "<title>19970</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1289.5,-125 1020.5,-125 1020.5,-72 1289.5,-72 1289.5,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"1155\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">type#1@List_of_cities_in_Argentina &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"1155\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 1101</text>\n",
       "<text text-anchor=\"middle\" x=\"1155\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [10, 29, 117, 945]</text>\n",
       "</g>\n",
       "<!-- 16466&#45;&gt;19970 -->\n",
       "<g id=\"edge22\" class=\"edge\">\n",
       "<title>16466&#45;&gt;19970</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1150.06,-160.87C1150.7,-152.89 1151.43,-143.93 1152.11,-135.38\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1155.61,-135.5 1152.93,-125.25 1148.64,-134.94 1155.61,-135.5\"/>\n",
       "</g>\n",
       "<!-- 16468 -->\n",
       "<g id=\"node21\" class=\"node\">\n",
       "<title>16468</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"910,-36 856,-36 856,0 910,0 910,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"883\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 16467&#45;&gt;16468 -->\n",
       "<g id=\"edge20\" class=\"edge\">\n",
       "<title>16467&#45;&gt;16468</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M899.14,-71.8C896.61,-63.53 893.81,-54.35 891.25,-46\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"894.57,-44.87 888.3,-36.33 887.88,-46.92 894.57,-44.87\"/>\n",
       "</g>\n",
       "<!-- 18745 -->\n",
       "<g id=\"node22\" class=\"node\">\n",
       "<title>18745</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"982,-36 928,-36 928,0 982,0 982,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"955\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 16467&#45;&gt;18745 -->\n",
       "<g id=\"edge21\" class=\"edge\">\n",
       "<title>16467&#45;&gt;18745</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M922.72,-71.8C927.99,-63.17 933.88,-53.54 939.16,-44.9\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"942.17,-46.69 944.4,-36.33 936.2,-43.04 942.17,-46.69\"/>\n",
       "</g>\n",
       "<!-- 19971 -->\n",
       "<g id=\"node24\" class=\"node\">\n",
       "<title>19971</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1146,-36 1092,-36 1092,0 1146,0 1146,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"1119\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 19970&#45;&gt;19971 -->\n",
       "<g id=\"edge23\" class=\"edge\">\n",
       "<title>19970&#45;&gt;19971</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1143.21,-71.8C1139.34,-63.35 1135.03,-53.95 1131.13,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1134.3,-43.96 1126.95,-36.33 1127.93,-46.88 1134.3,-43.96\"/>\n",
       "</g>\n",
       "<!-- 20218 -->\n",
       "<g id=\"node25\" class=\"node\">\n",
       "<title>20218</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1218,-36 1164,-36 1164,0 1218,0 1218,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"1191\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 19970&#45;&gt;20218 -->\n",
       "<g id=\"edge24\" class=\"edge\">\n",
       "<title>19970&#45;&gt;20218</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1166.79,-71.8C1170.66,-63.35 1174.97,-53.95 1178.87,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1182.07,-46.88 1183.05,-36.33 1175.7,-43.96 1182.07,-46.88\"/>\n",
       "</g>\n",
       "<!-- 20286 -->\n",
       "<g id=\"node27\" class=\"node\">\n",
       "<title>20286</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1486,-125 1308,-125 1308,-72 1486,-72 1486,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"1397\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">areaTotal#1 &lt;= 3950000.0</text>\n",
       "<text text-anchor=\"middle\" x=\"1397\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 2061</text>\n",
       "<text text-anchor=\"middle\" x=\"1397\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [628, 639, 549, 245]</text>\n",
       "</g>\n",
       "<!-- 20285&#45;&gt;20286 -->\n",
       "<g id=\"edge26\" class=\"edge\">\n",
       "<title>20285&#45;&gt;20286</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1402.64,-160.87C1401.91,-152.89 1401.09,-143.93 1400.3,-135.38\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1403.77,-134.89 1399.37,-125.25 1396.8,-135.53 1403.77,-134.89\"/>\n",
       "</g>\n",
       "<!-- 21763 -->\n",
       "<g id=\"node30\" class=\"node\">\n",
       "<title>21763</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1680,-125 1504,-125 1504,-72 1680,-72 1680,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"1592\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">utcOffset#1@THER &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"1592\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 567</text>\n",
       "<text text-anchor=\"middle\" x=\"1592\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [42, 80, 181, 264]</text>\n",
       "</g>\n",
       "<!-- 20285&#45;&gt;21763 -->\n",
       "<g id=\"edge29\" class=\"edge\">\n",
       "<title>20285&#45;&gt;21763</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1459.84,-160.98C1481.15,-151.07 1505.72,-139.64 1527.84,-129.35\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1529.57,-132.41 1537.16,-125.01 1526.62,-126.06 1529.57,-132.41\"/>\n",
       "</g>\n",
       "<!-- 20287 -->\n",
       "<g id=\"node28\" class=\"node\">\n",
       "<title>20287</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1377,-36 1323,-36 1323,0 1377,0 1377,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"1350\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 20286&#45;&gt;20287 -->\n",
       "<g id=\"edge27\" class=\"edge\">\n",
       "<title>20286&#45;&gt;20287</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1381.61,-71.8C1376.5,-63.26 1370.8,-53.75 1365.67,-45.17\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1368.52,-43.11 1360.38,-36.33 1362.51,-46.71 1368.52,-43.11\"/>\n",
       "</g>\n",
       "<!-- 21228 -->\n",
       "<g id=\"node29\" class=\"node\">\n",
       "<title>21228</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1449,-36 1395,-36 1395,0 1449,0 1449,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"1422\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 20286&#45;&gt;21228 -->\n",
       "<g id=\"edge28\" class=\"edge\">\n",
       "<title>20286&#45;&gt;21228</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1405.19,-71.8C1407.82,-63.53 1410.74,-54.35 1413.4,-46\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1416.78,-46.92 1416.48,-36.33 1410.11,-44.8 1416.78,-46.92\"/>\n",
       "</g>\n",
       "<!-- 21764 -->\n",
       "<g id=\"node31\" class=\"node\">\n",
       "<title>21764</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1583,-36 1529,-36 1529,0 1583,0 1583,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"1556\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 21763&#45;&gt;21764 -->\n",
       "<g id=\"edge30\" class=\"edge\">\n",
       "<title>21763&#45;&gt;21764</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1580.21,-71.8C1576.34,-63.35 1572.03,-53.95 1568.13,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1571.3,-43.96 1563.95,-36.33 1564.93,-46.88 1571.3,-43.96\"/>\n",
       "</g>\n",
       "<!-- 21899 -->\n",
       "<g id=\"node32\" class=\"node\">\n",
       "<title>21899</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1655,-36 1601,-36 1601,0 1655,0 1655,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"1628\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 21763&#45;&gt;21899 -->\n",
       "<g id=\"edge31\" class=\"edge\">\n",
       "<title>21763&#45;&gt;21899</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1603.79,-71.8C1607.66,-63.35 1611.97,-53.95 1615.87,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1619.07,-46.88 1620.05,-36.33 1612.7,-43.96 1619.07,-46.88\"/>\n",
       "</g>\n",
       "<!-- 22043 -->\n",
       "<g id=\"node34\" class=\"node\">\n",
       "<title>22043</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2235,-303 2041,-303 2041,-250 2235,-250 2235,-303\"/>\n",
       "<text text-anchor=\"middle\" x=\"2138\" y=\"-287.8\" font-family=\"Times,serif\" font-size=\"14.00\">birthPlace?inv#count &lt;= 1.389</text>\n",
       "<text text-anchor=\"middle\" x=\"2138\" y=\"-272.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 4618</text>\n",
       "<text text-anchor=\"middle\" x=\"2138\" y=\"-257.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [180, 572, 1772, 2094]</text>\n",
       "</g>\n",
       "<!-- 22042&#45;&gt;22043 -->\n",
       "<g id=\"edge33\" class=\"edge\">\n",
       "<title>22042&#45;&gt;22043</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2138,-338.87C2138,-330.89 2138,-321.93 2138,-313.38\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2141.5,-313.25 2138,-303.25 2134.5,-313.25 2141.5,-313.25\"/>\n",
       "</g>\n",
       "<!-- 23872 -->\n",
       "<g id=\"node49\" class=\"node\">\n",
       "<title>23872</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2831.5,-303 2580.5,-303 2580.5,-250 2831.5,-250 2831.5,-303\"/>\n",
       "<text text-anchor=\"middle\" x=\"2706\" y=\"-287.8\" font-family=\"Times,serif\" font-size=\"14.00\">22&#45;rdf&#45;syntax&#45;ns#type#1@THER &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"2706\" y=\"-272.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 2645</text>\n",
       "<text text-anchor=\"middle\" x=\"2706\" y=\"-257.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [18, 55, 310, 2262]</text>\n",
       "</g>\n",
       "<!-- 22042&#45;&gt;23872 -->\n",
       "<g id=\"edge48\" class=\"edge\">\n",
       "<title>22042&#45;&gt;23872</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2233.71,-349.84C2326.06,-335.7 2466.74,-314.15 2570.56,-298.25\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2571.13,-301.7 2580.49,-296.72 2570.07,-294.78 2571.13,-301.7\"/>\n",
       "</g>\n",
       "<!-- 22044 -->\n",
       "<g id=\"node35\" class=\"node\">\n",
       "<title>22044</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2132.5,-214 1923.5,-214 1923.5,-161 2132.5,-161 2132.5,-214\"/>\n",
       "<text text-anchor=\"middle\" x=\"2028\" y=\"-198.8\" font-family=\"Times,serif\" font-size=\"14.00\">country#1@United_States &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"2028\" y=\"-183.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 3862</text>\n",
       "<text text-anchor=\"middle\" x=\"2028\" y=\"-168.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [179, 563, 1667, 1453]</text>\n",
       "</g>\n",
       "<!-- 22043&#45;&gt;22044 -->\n",
       "<g id=\"edge34\" class=\"edge\">\n",
       "<title>22043&#45;&gt;22044</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2105.59,-249.87C2093.95,-240.66 2080.67,-230.15 2068.42,-220.47\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2070.57,-217.71 2060.56,-214.25 2066.23,-223.2 2070.57,-217.71\"/>\n",
       "</g>\n",
       "<!-- 23695 -->\n",
       "<g id=\"node42\" class=\"node\">\n",
       "<title>23695</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2345,-214 2151,-214 2151,-161 2345,-161 2345,-214\"/>\n",
       "<text text-anchor=\"middle\" x=\"2248\" y=\"-198.8\" font-family=\"Times,serif\" font-size=\"14.00\">birthPlace?inv#count &lt;= 1.574</text>\n",
       "<text text-anchor=\"middle\" x=\"2248\" y=\"-183.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 756</text>\n",
       "<text text-anchor=\"middle\" x=\"2248\" y=\"-168.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [1, 9, 105, 641]</text>\n",
       "</g>\n",
       "<!-- 22043&#45;&gt;23695 -->\n",
       "<g id=\"edge41\" class=\"edge\">\n",
       "<title>22043&#45;&gt;23695</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2170.41,-249.87C2182.05,-240.66 2195.33,-230.15 2207.58,-220.47\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2209.77,-223.2 2215.44,-214.25 2205.43,-217.71 2209.77,-223.2\"/>\n",
       "</g>\n",
       "<!-- 22045 -->\n",
       "<g id=\"node36\" class=\"node\">\n",
       "<title>22045</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1927.5,-125 1698.5,-125 1698.5,-72 1927.5,-72 1927.5,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"1813\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">country#1@United_Kingdom &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"1813\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 2537</text>\n",
       "<text text-anchor=\"middle\" x=\"1813\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [86, 238, 977, 1236]</text>\n",
       "</g>\n",
       "<!-- 22044&#45;&gt;22045 -->\n",
       "<g id=\"edge35\" class=\"edge\">\n",
       "<title>22044&#45;&gt;22045</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1964.95,-160.98C1940.01,-150.89 1911.19,-139.23 1885.4,-128.8\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1886.64,-125.52 1876.05,-125.01 1884.01,-132.01 1886.64,-125.52\"/>\n",
       "</g>\n",
       "<!-- 23308 -->\n",
       "<g id=\"node39\" class=\"node\">\n",
       "<title>23308</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2118.5,-125 1945.5,-125 1945.5,-72 2118.5,-72 2118.5,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"2032\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">areaTotal#1 &lt;= 13075000.0</text>\n",
       "<text text-anchor=\"middle\" x=\"2032\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 1325</text>\n",
       "<text text-anchor=\"middle\" x=\"2032\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [93, 325, 690, 217]</text>\n",
       "</g>\n",
       "<!-- 22044&#45;&gt;23308 -->\n",
       "<g id=\"edge38\" class=\"edge\">\n",
       "<title>22044&#45;&gt;23308</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2029.18,-160.87C2029.55,-152.89 2029.96,-143.93 2030.35,-135.38\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2033.85,-135.4 2030.82,-125.25 2026.86,-135.08 2033.85,-135.4\"/>\n",
       "</g>\n",
       "<!-- 22046 -->\n",
       "<g id=\"node37\" class=\"node\">\n",
       "<title>22046</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1804,-36 1750,-36 1750,0 1804,0 1804,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"1777\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 22045&#45;&gt;22046 -->\n",
       "<g id=\"edge36\" class=\"edge\">\n",
       "<title>22045&#45;&gt;22046</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1801.21,-71.8C1797.34,-63.35 1793.03,-53.95 1789.13,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1792.3,-43.96 1784.95,-36.33 1785.93,-46.88 1792.3,-43.96\"/>\n",
       "</g>\n",
       "<!-- 22901 -->\n",
       "<g id=\"node38\" class=\"node\">\n",
       "<title>22901</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"1876,-36 1822,-36 1822,0 1876,0 1876,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"1849\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 22045&#45;&gt;22901 -->\n",
       "<g id=\"edge37\" class=\"edge\">\n",
       "<title>22045&#45;&gt;22901</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M1824.79,-71.8C1828.66,-63.35 1832.97,-53.95 1836.87,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"1840.07,-46.88 1841.05,-36.33 1833.7,-43.96 1840.07,-46.88\"/>\n",
       "</g>\n",
       "<!-- 23309 -->\n",
       "<g id=\"node40\" class=\"node\">\n",
       "<title>23309</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2023,-36 1969,-36 1969,0 2023,0 2023,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"1996\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 23308&#45;&gt;23309 -->\n",
       "<g id=\"edge39\" class=\"edge\">\n",
       "<title>23308&#45;&gt;23309</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2020.21,-71.8C2016.34,-63.35 2012.03,-53.95 2008.13,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2011.3,-43.96 2003.95,-36.33 2004.93,-46.88 2011.3,-43.96\"/>\n",
       "</g>\n",
       "<!-- 23558 -->\n",
       "<g id=\"node41\" class=\"node\">\n",
       "<title>23558</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2095,-36 2041,-36 2041,0 2095,0 2095,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"2068\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 23308&#45;&gt;23558 -->\n",
       "<g id=\"edge40\" class=\"edge\">\n",
       "<title>23308&#45;&gt;23558</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2043.79,-71.8C2047.66,-63.35 2051.97,-53.95 2055.87,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2059.07,-46.88 2060.05,-36.33 2052.7,-43.96 2059.07,-46.88\"/>\n",
       "</g>\n",
       "<!-- 23696 -->\n",
       "<g id=\"node43\" class=\"node\">\n",
       "<title>23696</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2351,-125 2137,-125 2137,-72 2351,-72 2351,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"2244\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">routeJunction?inv#count &lt;= 0.151</text>\n",
       "<text text-anchor=\"middle\" x=\"2244\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 332</text>\n",
       "<text text-anchor=\"middle\" x=\"2244\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [1, 6, 83, 242]</text>\n",
       "</g>\n",
       "<!-- 23695&#45;&gt;23696 -->\n",
       "<g id=\"edge42\" class=\"edge\">\n",
       "<title>23695&#45;&gt;23696</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2246.82,-160.87C2246.45,-152.89 2246.04,-143.93 2245.65,-135.38\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2249.14,-135.08 2245.18,-125.25 2242.15,-135.4 2249.14,-135.08\"/>\n",
       "</g>\n",
       "<!-- 23807 -->\n",
       "<g id=\"node46\" class=\"node\">\n",
       "<title>23807</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2513,-125 2369,-125 2369,-72 2513,-72 2513,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"2441\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">rel#count &lt;= 1.789</text>\n",
       "<text text-anchor=\"middle\" x=\"2441\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 424</text>\n",
       "<text text-anchor=\"middle\" x=\"2441\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [0, 3, 22, 399]</text>\n",
       "</g>\n",
       "<!-- 23695&#45;&gt;23807 -->\n",
       "<g id=\"edge45\" class=\"edge\">\n",
       "<title>23695&#45;&gt;23807</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2304.6,-160.98C2326.7,-151.03 2352.18,-139.54 2375.09,-129.21\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2376.72,-132.31 2384.4,-125.01 2373.84,-125.93 2376.72,-132.31\"/>\n",
       "</g>\n",
       "<!-- 23697 -->\n",
       "<g id=\"node44\" class=\"node\">\n",
       "<title>23697</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2235,-36 2181,-36 2181,0 2235,0 2235,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"2208\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 23696&#45;&gt;23697 -->\n",
       "<g id=\"edge43\" class=\"edge\">\n",
       "<title>23696&#45;&gt;23697</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2232.21,-71.8C2228.34,-63.35 2224.03,-53.95 2220.13,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2223.3,-43.96 2215.95,-36.33 2216.93,-46.88 2223.3,-43.96\"/>\n",
       "</g>\n",
       "<!-- 23770 -->\n",
       "<g id=\"node45\" class=\"node\">\n",
       "<title>23770</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2307,-36 2253,-36 2253,0 2307,0 2307,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"2280\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 23696&#45;&gt;23770 -->\n",
       "<g id=\"edge44\" class=\"edge\">\n",
       "<title>23696&#45;&gt;23770</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2255.79,-71.8C2259.66,-63.35 2263.97,-53.95 2267.87,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2271.07,-46.88 2272.05,-36.33 2264.7,-43.96 2271.07,-46.88\"/>\n",
       "</g>\n",
       "<!-- 23808 -->\n",
       "<g id=\"node47\" class=\"node\">\n",
       "<title>23808</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2432,-36 2378,-36 2378,0 2432,0 2432,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"2405\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 23807&#45;&gt;23808 -->\n",
       "<g id=\"edge46\" class=\"edge\">\n",
       "<title>23807&#45;&gt;23808</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2429.21,-71.8C2425.34,-63.35 2421.03,-53.95 2417.13,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2420.3,-43.96 2412.95,-36.33 2413.93,-46.88 2420.3,-43.96\"/>\n",
       "</g>\n",
       "<!-- 23813 -->\n",
       "<g id=\"node48\" class=\"node\">\n",
       "<title>23813</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2504,-36 2450,-36 2450,0 2504,0 2504,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"2477\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 23807&#45;&gt;23813 -->\n",
       "<g id=\"edge47\" class=\"edge\">\n",
       "<title>23807&#45;&gt;23813</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2452.79,-71.8C2456.66,-63.35 2460.97,-53.95 2464.87,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2468.07,-46.88 2469.05,-36.33 2461.7,-43.96 2468.07,-46.88\"/>\n",
       "</g>\n",
       "<!-- 23873 -->\n",
       "<g id=\"node50\" class=\"node\">\n",
       "<title>23873</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2803,-214 2609,-214 2609,-161 2803,-161 2803,-214\"/>\n",
       "<text text-anchor=\"middle\" x=\"2706\" y=\"-198.8\" font-family=\"Times,serif\" font-size=\"14.00\">birthPlace?inv#count &lt;= 1.161</text>\n",
       "<text text-anchor=\"middle\" x=\"2706\" y=\"-183.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 2134</text>\n",
       "<text text-anchor=\"middle\" x=\"2706\" y=\"-168.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [5, 20, 155, 1954]</text>\n",
       "</g>\n",
       "<!-- 23872&#45;&gt;23873 -->\n",
       "<g id=\"edge49\" class=\"edge\">\n",
       "<title>23872&#45;&gt;23873</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2706,-249.87C2706,-241.89 2706,-232.93 2706,-224.38\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2709.5,-224.25 2706,-214.25 2702.5,-224.25 2709.5,-224.25\"/>\n",
       "</g>\n",
       "<!-- 24132 -->\n",
       "<g id=\"node57\" class=\"node\">\n",
       "<title>24132</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"3071,-214 2907,-214 2907,-161 3071,-161 3071,-214\"/>\n",
       "<text text-anchor=\"middle\" x=\"2989\" y=\"-198.8\" font-family=\"Times,serif\" font-size=\"14.00\">utcOffset#2@4 &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"2989\" y=\"-183.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 511</text>\n",
       "<text text-anchor=\"middle\" x=\"2989\" y=\"-168.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [13, 35, 155, 308]</text>\n",
       "</g>\n",
       "<!-- 23872&#45;&gt;24132 -->\n",
       "<g id=\"edge56\" class=\"edge\">\n",
       "<title>23872&#45;&gt;24132</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2789,-249.98C2822.99,-239.54 2862.44,-227.41 2897.29,-216.7\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2898.45,-220 2906.98,-213.71 2896.39,-213.31 2898.45,-220\"/>\n",
       "</g>\n",
       "<!-- 23874 -->\n",
       "<g id=\"node51\" class=\"node\">\n",
       "<title>23874</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2707,-125 2531,-125 2531,-72 2707,-72 2707,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"2619\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">utcOffset#2@THER &lt;= 0.5</text>\n",
       "<text text-anchor=\"middle\" x=\"2619\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 891</text>\n",
       "<text text-anchor=\"middle\" x=\"2619\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [3, 20, 127, 741]</text>\n",
       "</g>\n",
       "<!-- 23873&#45;&gt;23874 -->\n",
       "<g id=\"edge50\" class=\"edge\">\n",
       "<title>23873&#45;&gt;23874</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2680.37,-160.87C2671.51,-152.01 2661.45,-141.95 2652.08,-132.58\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2654.3,-129.85 2644.75,-125.25 2649.35,-134.8 2654.3,-129.85\"/>\n",
       "</g>\n",
       "<!-- 24071 -->\n",
       "<g id=\"node54\" class=\"node\">\n",
       "<title>24071</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2898.5,-125 2725.5,-125 2725.5,-72 2898.5,-72 2898.5,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"2812\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">areaTotal#1 &lt;= 57828368.0</text>\n",
       "<text text-anchor=\"middle\" x=\"2812\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 1243</text>\n",
       "<text text-anchor=\"middle\" x=\"2812\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [2, 0, 28, 1213]</text>\n",
       "</g>\n",
       "<!-- 23873&#45;&gt;24071 -->\n",
       "<g id=\"edge53\" class=\"edge\">\n",
       "<title>23873&#45;&gt;24071</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2737.23,-160.87C2748.34,-151.75 2761.01,-141.35 2772.71,-131.75\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2775.12,-134.3 2780.63,-125.25 2770.68,-128.89 2775.12,-134.3\"/>\n",
       "</g>\n",
       "<!-- 23875 -->\n",
       "<g id=\"node52\" class=\"node\">\n",
       "<title>23875</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2611,-36 2557,-36 2557,0 2611,0 2611,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"2584\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 23874&#45;&gt;23875 -->\n",
       "<g id=\"edge51\" class=\"edge\">\n",
       "<title>23874&#45;&gt;23875</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2607.54,-71.8C2603.81,-63.44 2599.67,-54.15 2595.91,-45.72\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2599,-44.04 2591.73,-36.33 2592.6,-46.89 2599,-44.04\"/>\n",
       "</g>\n",
       "<!-- 23978 -->\n",
       "<g id=\"node53\" class=\"node\">\n",
       "<title>23978</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2683,-36 2629,-36 2629,0 2683,0 2683,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"2656\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 23874&#45;&gt;23978 -->\n",
       "<g id=\"edge52\" class=\"edge\">\n",
       "<title>23874&#45;&gt;23978</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2631.12,-71.8C2635.1,-63.35 2639.53,-53.95 2643.53,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2646.73,-46.87 2647.83,-36.33 2640.4,-43.88 2646.73,-46.87\"/>\n",
       "</g>\n",
       "<!-- 24072 -->\n",
       "<g id=\"node55\" class=\"node\">\n",
       "<title>24072</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2803,-36 2749,-36 2749,0 2803,0 2803,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"2776\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 24071&#45;&gt;24072 -->\n",
       "<g id=\"edge54\" class=\"edge\">\n",
       "<title>24071&#45;&gt;24072</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2800.21,-71.8C2796.34,-63.35 2792.03,-53.95 2788.13,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2791.3,-43.96 2783.95,-36.33 2784.93,-46.88 2791.3,-43.96\"/>\n",
       "</g>\n",
       "<!-- 24093 -->\n",
       "<g id=\"node56\" class=\"node\">\n",
       "<title>24093</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"2875,-36 2821,-36 2821,0 2875,0 2875,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"2848\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 24071&#45;&gt;24093 -->\n",
       "<g id=\"edge55\" class=\"edge\">\n",
       "<title>24071&#45;&gt;24093</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2823.79,-71.8C2827.66,-63.35 2831.97,-53.95 2835.87,-45.45\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2839.07,-46.88 2840.05,-36.33 2832.7,-43.96 2839.07,-46.88\"/>\n",
       "</g>\n",
       "<!-- 24133 -->\n",
       "<g id=\"node58\" class=\"node\">\n",
       "<title>24133</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"3061,-125 2917,-125 2917,-72 3061,-72 3061,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"2989\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">rel#count &lt;= 1.243</text>\n",
       "<text text-anchor=\"middle\" x=\"2989\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 333</text>\n",
       "<text text-anchor=\"middle\" x=\"2989\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [1, 7, 80, 245]</text>\n",
       "</g>\n",
       "<!-- 24132&#45;&gt;24133 -->\n",
       "<g id=\"edge57\" class=\"edge\">\n",
       "<title>24132&#45;&gt;24133</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2989,-160.87C2989,-152.89 2989,-143.93 2989,-135.38\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2992.5,-135.25 2989,-125.25 2985.5,-135.25 2992.5,-135.25\"/>\n",
       "</g>\n",
       "<!-- 24262 -->\n",
       "<g id=\"node61\" class=\"node\">\n",
       "<title>24262</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"3285,-125 3079,-125 3079,-72 3285,-72 3285,-125\"/>\n",
       "<text text-anchor=\"middle\" x=\"3182\" y=\"-109.8\" font-family=\"Times,serif\" font-size=\"14.00\">populationDensity#1 &lt;= 383.465</text>\n",
       "<text text-anchor=\"middle\" x=\"3182\" y=\"-94.8\" font-family=\"Times,serif\" font-size=\"14.00\">samples = 178</text>\n",
       "<text text-anchor=\"middle\" x=\"3182\" y=\"-79.8\" font-family=\"Times,serif\" font-size=\"14.00\">value = [12, 28, 75, 63]</text>\n",
       "</g>\n",
       "<!-- 24132&#45;&gt;24262 -->\n",
       "<g id=\"edge60\" class=\"edge\">\n",
       "<title>24132&#45;&gt;24262</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M3045.6,-160.98C3067.7,-151.03 3093.18,-139.54 3116.09,-129.21\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"3117.72,-132.31 3125.4,-125.01 3114.84,-125.93 3117.72,-132.31\"/>\n",
       "</g>\n",
       "<!-- 24134 -->\n",
       "<g id=\"node59\" class=\"node\">\n",
       "<title>24134</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"3002,-36 2948,-36 2948,0 3002,0 3002,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"2975\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 24133&#45;&gt;24134 -->\n",
       "<g id=\"edge58\" class=\"edge\">\n",
       "<title>24133&#45;&gt;24134</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M2984.42,-71.8C2982.96,-63.62 2981.34,-54.55 2979.86,-46.27\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"2983.29,-45.56 2978.09,-36.33 2976.4,-46.79 2983.29,-45.56\"/>\n",
       "</g>\n",
       "<!-- 24149 -->\n",
       "<g id=\"node60\" class=\"node\">\n",
       "<title>24149</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"3074,-36 3020,-36 3020,0 3074,0 3074,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"3047\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 24133&#45;&gt;24149 -->\n",
       "<g id=\"edge59\" class=\"edge\">\n",
       "<title>24133&#45;&gt;24149</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M3007.99,-71.8C3014.43,-63.08 3021.63,-53.34 3028.07,-44.63\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"3031.07,-46.45 3034.2,-36.33 3025.44,-42.29 3031.07,-46.45\"/>\n",
       "</g>\n",
       "<!-- 24263 -->\n",
       "<g id=\"node62\" class=\"node\">\n",
       "<title>24263</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"3151,-36 3097,-36 3097,0 3151,0 3151,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"3124\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 24262&#45;&gt;24263 -->\n",
       "<g id=\"edge61\" class=\"edge\">\n",
       "<title>24262&#45;&gt;24263</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M3163.01,-71.8C3156.57,-63.08 3149.37,-53.34 3142.93,-44.63\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"3145.56,-42.29 3136.8,-36.33 3139.93,-46.45 3145.56,-42.29\"/>\n",
       "</g>\n",
       "<!-- 24354 -->\n",
       "<g id=\"node63\" class=\"node\">\n",
       "<title>24354</title>\n",
       "<polygon fill=\"none\" stroke=\"black\" points=\"3223,-36 3169,-36 3169,0 3223,0 3223,-36\"/>\n",
       "<text text-anchor=\"middle\" x=\"3196\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">(...)</text>\n",
       "</g>\n",
       "<!-- 24262&#45;&gt;24354 -->\n",
       "<g id=\"edge62\" class=\"edge\">\n",
       "<title>24262&#45;&gt;24354</title>\n",
       "<path fill=\"none\" stroke=\"black\" d=\"M3186.58,-71.8C3188.04,-63.62 3189.66,-54.55 3191.14,-46.27\"/>\n",
       "<polygon fill=\"black\" stroke=\"black\" points=\"3194.6,-46.79 3192.91,-36.33 3187.71,-45.56 3194.6,-46.79\"/>\n",
       "</g>\n",
       "</g>\n",
       "</svg>\n"
      ],
      "text/plain": [
       "<graphviz.files.Source at 0x7f92c4a947d0>"
      ]
     },
     "execution_count": 29,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# CELL 29\n",
    "import numpy as np\n",
    "import pickle\n",
    "\n",
    "with open(\"ch6_cell27_splits.pk\", \"rb\") as pkl:\n",
    "    segments_at = pickle.load(pkl)\n",
    "\n",
    "data = list()\n",
    "header = None\n",
    "with open(\"ch6_cell28_dev_feat1_filtered.tsv\") as feats:\n",
    "    header = next(feats)\n",
    "    header = header.strip().split(\"\\t\")\n",
    "    header.pop(0) # name\n",
    "    header.pop() # pop\n",
    "    for line in feats:\n",
    "        fields = line.strip().split(\"\\t\")\n",
    "        logpop = float(fields[-1])\n",
    "        name = fields[0]\n",
    "        feats = list(map(float,fields[1:-1]))\n",
    "        row = (feats, logpop, name)\n",
    "        data.append(row)\n",
    "\n",
    "xdata = np.array(list(map(lambda t:t[0], data)))\n",
    "ydata = np.array(list(map(lambda t:t[1], data)))\n",
    "\n",
    "ydata = cell28_adjudicate(ydata, segments_at[2])\n",
    "print_names = list()\n",
    "for name in header:\n",
    "    name = name.replace('\"','')\n",
    "    if '@' in name:\n",
    "        pos = name.find('@')\n",
    "        rel, val = name[0:pos], name[pos+2:]\n",
    "        rel = rel.replace('<','').replace('>','').split(\"/\")[-1]\n",
    "        val = val.replace('<','').replace('>','').split(\"/\")[-1]\n",
    "        name = \"{}@{}\".format(rel, val)\n",
    "    else:\n",
    "        name = name.replace('<','').replace('>','').split(\"/\")[-1]\n",
    "    print_names.append(name)   \n",
    "\n",
    "from sklearn import tree\n",
    "clf = tree.DecisionTreeClassifier()\n",
    "clf.fit(xdata, ydata)\n",
    "\n",
    "import graphviz\n",
    "dot_data = tree.export_graphviz(clf, out_file=None, max_depth=4,\n",
    "                               impurity=False, proportion=False,\n",
    "                               feature_names=print_names)\n",
    "graph = graphviz.Source(dot_data)\n",
    "graph.render(\"ch6_cell29_tree.dot\")\n",
    "graph"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Second Featurization\n",
    "\n",
    "From the previous analysis I decided to a computable feature using $area \\times density$ and to perform an out-of-fold target-rate encoding in Cell #30, producing \"ch6_cell30_dev_feats2_train.tsv\" and \"ch6_cell30_dev_feats2_test.tsv\".\n",
    "\n",
    "Due to the impossiblity of doing TRE and normalization, this new feature set is already normalized.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Training on 35,971 cities\n",
      "RMSE 0.3298424854334205\n",
      "Writing files for error analysis\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x7f93104c7f10>"
      ]
     },
     "execution_count": 30,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x360 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 30\n",
    "import random\n",
    "from functools import reduce\n",
    "from sklearn.svm import SVR\n",
    "import numpy as np\n",
    "from collections import OrderedDict\n",
    "\n",
    "rand = random.Random(42)\n",
    "train_data = list()\n",
    "test_data  = list()\n",
    "header = None\n",
    "with open(\"ch6_cell28_dev_feat1_filtered.tsv\") as feats:\n",
    "    header = next(feats)\n",
    "    header = header.strip().split(\"\\t\")\n",
    "    header.pop(0) # name\n",
    "    header.pop() # population\n",
    "    for line in feats:\n",
    "        fields = line.strip().split(\"\\t\")\n",
    "        logpop = float(fields[-1])\n",
    "        name = fields[0]\n",
    "        feats = list(map(float,fields[1:-1]))\n",
    "        row = (feats, logpop, name)\n",
    "        if rand.random() < 0.2:\n",
    "            test_data.append(row) \n",
    "        else:\n",
    "            train_data.append(row)\n",
    "               \n",
    "test_data  = sorted(test_data, key=lambda t:t[1])\n",
    "test_names = list(map(lambda t:t[2], test_data))\n",
    "\n",
    "xtrain = np.array(list(map(lambda t:t[0], train_data)))\n",
    "ytrain = np.array(list(map(lambda t:t[1], train_data)))\n",
    "xtest  = np.array(list(map(lambda t:t[0], test_data)))\n",
    "ytest  = np.array(list(map(lambda t:t[1], test_data)))\n",
    "\n",
    "# add computed feature density area times\n",
    "approx_area_train = np.zeros( (xtrain.shape[0],) )\n",
    "approx_area_test  = np.zeros( (xtest.shape[0],)  )\n",
    "#BUG \"area#1\" is not included here. It should\n",
    "for idx, name in enumerate(header):\n",
    "    if \"#count\" not in name and (\"areaLand#\" in name or \"areaTotal#\" in name):\n",
    "        approx_area_train = np.max( [ approx_area_train, xtrain[:,idx] ], axis=0)\n",
    "        approx_area_test =  np.max( [ approx_area_test, xtest[:,idx] ], axis=0)\n",
    "density_idx = next(filter(lambda t:\"populationDensity#1\" in t[1], \n",
    "                     enumerate(header)))[0]\n",
    "computed_value_train = approx_area_train * xtrain[:,density_idx]\n",
    "computed_value_test  = approx_area_test  * xtest[:, density_idx]\n",
    "computed_defined_train = np.array( list(map(lambda x: 1.0 if x>0 else 0.0, \n",
    "                                            computed_value_train)) )\n",
    "computed_defined_test  = np.array( list(map(lambda x: 1.0 if x>0 else 0.0, \n",
    "                                           computed_value_test)) )\n",
    "computed_value_train = np.log(computed_value_train + 1)\n",
    "computed_value_test  = np.log(computed_value_test  + 1)\n",
    "header.append(\"computed#defined\")\n",
    "header.append(\"computed#value\")\n",
    "xtrain = np.append(xtrain, computed_defined_train.reshape( (xtrain.shape[0],1) ), axis=1)\n",
    "xtrain = np.append(xtrain, computed_value_train.reshape(   (xtrain.shape[0],1) ), axis=1)\n",
    "xtest  = np.append(xtest,  computed_defined_test.reshape(  (xtest.shape[0],1) ),  axis=1)\n",
    "xtest  = np.append(xtest,  computed_value_test.reshape(    (xtest.shape[0],1) ),  axis=1)\n",
    "\n",
    "# out-of-fold target encoding for tsv\n",
    "xtrain_non_tre = xtrain.copy()\n",
    "xtest_non_tre  = xtest.copy()\n",
    "\n",
    "def doTRE(xtrain, ytrain, xtest, ytest):\n",
    "    PARAM_FOLDS = 10\n",
    "    PARAM_ADD_ORIGINAL_COLS = True\n",
    "    PARAM_SINGLE_COL_TRE = True\n",
    "    folds = list(map(lambda x:rand.randint(0, PARAM_FOLDS - 1), range(xtrain.shape[0])))\n",
    "    # value to use when there is no evidence available\n",
    "    imputed_tre = sorted(list(ytrain))[int(ytrain.shape[0]/2)]\n",
    "    \n",
    "    # gather the TRE cols\n",
    "    tre_cols = list()\n",
    "    tre_col_names = list()\n",
    "    cols_to_tre_cols = dict()\n",
    "    cats_for_attrib = OrderedDict() # name to list of idx\n",
    "    for idx, name in enumerate(header):\n",
    "        if '@' in name: # categorical value\n",
    "            tre_idx = len(tre_cols)\n",
    "            cols_to_tre_cols[idx] = tre_idx\n",
    "            tre_cols.append(idx)\n",
    "            tre_col_names.append(name + '+TRE')\n",
    "            attrib = name[0:name.index('@')]\n",
    "            if attrib not in cats_for_attrib:\n",
    "                cats_for_attrib[attrib] = []\n",
    "            cats_for_attrib[attrib].append( (idx, tre_idx) )\n",
    "    tre_train = np.zeros( (xtrain.shape[0], len(tre_cols)) )\n",
    "    tre_test  = np.zeros( (xtest.shape[0],  len(tre_cols)) )\n",
    "    \n",
    "    saved_full_TRE = OrderedDict()\n",
    "    \n",
    "    for idx, name in enumerate(header):\n",
    "        if '@' in name: # categorical value\n",
    "            tre_col = cols_to_tre_cols[idx]\n",
    "            \n",
    "            fold_val_pops = { fold: dict() for fold in range(PARAM_FOLDS) }\n",
    "            # gather the pop values for the features in all folds\n",
    "            for row in range(xtrain.shape[0]):\n",
    "                value = xtrain[row, idx]\n",
    "                fold = folds[row]\n",
    "                fold_val_pops[fold][value] = fold_val_pops[fold].get(value, []) + [ytrain[row]]\n",
    "            all_values = set(xtrain[:,idx]).union(set(xtest[:,idx]))\n",
    "            # compute TREs\n",
    "            fold_TREs = list()\n",
    "            for fold in range(PARAM_FOLDS):\n",
    "                val_pops = dict()\n",
    "                for other in range(PARAM_FOLDS):\n",
    "                    if other != fold:\n",
    "                        for value in fold_val_pops[other].keys():\n",
    "                            val_pops[value] = val_pops.get(value, []) + fold_val_pops[other][value]\n",
    "                fold_TRE = OrderedDict()\n",
    "                for value in all_values:\n",
    "                    val_pops[value] = sorted(val_pops.get(value, []))\n",
    "                    fold_TRE[value] = imputed_tre if len(val_pops[value]) == 0 else \\\n",
    "                      val_pops[value][int(len(val_pops[value])/2)]\n",
    "                fold_TREs.append(fold_TRE)\n",
    "            full_val_pops = dict()\n",
    "            for value in all_values:\n",
    "                full_val_pops[value] = sorted(reduce(lambda x, y: x + y, \n",
    "                                                  map(lambda f:f.get(value, []), fold_val_pops.values()), []))\n",
    "            full_TRE = OrderedDict()\n",
    "            for value in all_values:\n",
    "                full_TRE[value] = imputed_tre if len(full_val_pops[value]) == 0 \\\n",
    "                   else full_val_pops[value][int(len(full_val_pops[value])/2)]\n",
    "\n",
    "            # encode folds\n",
    "            for row in range(xtrain.shape[0]):\n",
    "                tre_train[row, tre_col] = fold_TREs[folds[row]][xtrain[row, idx]]\n",
    "            # encode test (full TRE)\n",
    "            for row in range(xtest.shape[0]):\n",
    "                tre_test[row, tre_col] = full_TRE[xtest[row, idx]]\n",
    "            saved_full_TRE[idx] = full_TRE\n",
    "            \n",
    "    if PARAM_SINGLE_COL_TRE:\n",
    "        single_tre_train = np.zeros( (xtrain.shape[0], len(cats_for_attrib)) )\n",
    "        single_tre_test  = np.zeros( (xtest.shape[0],  len(cats_for_attrib)) )\n",
    "        single_tre_col_names = list()\n",
    "        single_tre_count = 0\n",
    "        for attrib, idxs in cats_for_attrib.items():\n",
    "            single_tre_col_names.append(attrib + '#TRE')\n",
    "            for row in range(xtrain.shape[0]):\n",
    "                max_tre_idx = None\n",
    "                max_val = None\n",
    "                for idx, tre_idx in idxs:\n",
    "                    if max_tre_idx is None or xtrain[row,idx] > max_val:\n",
    "                        max_val = xtrain[row,idx]\n",
    "                        max_tre_idx = tre_idx\n",
    "                single_tre_train[row,single_tre_count] = tre_train[row,tre_idx]\n",
    "            for row in range(xtest.shape[0]):\n",
    "                max_tre_idx = None\n",
    "                max_val = None\n",
    "                for idx, tre_idx in idxs:\n",
    "                    if max_tre_idx is None or xtest[row,idx] > max_val:\n",
    "                        max_val = xtest[row,idx]\n",
    "                        max_tre_idx = tre_idx\n",
    "                single_tre_test[row,single_tre_count] = tre_test[row,tre_idx]\n",
    "            single_tre_count += 1\n",
    "        tre_train = single_tre_train\n",
    "        tre_test  = single_tre_test\n",
    "        tre_col_names = single_tre_col_names\n",
    "\n",
    "    if not PARAM_ADD_ORIGINAL_COLS:\n",
    "        # delete original columns\n",
    "        xtrain = np.delete(xtrain, tre_cols, 1)\n",
    "        xtest  = np.delete(xtest,  tre_cols, 1)\n",
    "        for idx in reversed(tre_cols):\n",
    "            del header[idx]\n",
    "            \n",
    "    # add tre columns\n",
    "    header.extend(tre_col_names)\n",
    "    xtrain = np.append(xtrain, tre_train, axis=1)\n",
    "    xtest = np.append(xtest, tre_test, axis=1)\n",
    "    \n",
    "    return header, xtrain, ytrain, xtest, ytest, saved_full_TRE\n",
    "\n",
    "PARAM_WRITE_UNSCALED = False\n",
    "if PARAM_WRITE_UNSCALED:\n",
    "    new_header, xtrain, ytrain, xtest, ytest, _ = doTRE(xtrain, ytrain, xtest, ytest)\n",
    "\n",
    "    # write new data\n",
    "    with open(\"ch6_cell30_dev_feats2_train_raw.tsv\", \"w\") as feats:\n",
    "        feats.write(\"\\t\".join([\"name\"] + new_header + [\"population\"])+\"\\n\")\n",
    "        arr, target, names = xtrain, ytrain, list(map(lambda t:t[2],train_data))\n",
    "        for row in range(arr.shape[0]):\n",
    "            feats.write(\"\\t\".join(list(map(lambda v:str(v),\n",
    "                                       [names[row]]+list(arr[row,:])+[target[row]]))) +\"\\n\")\n",
    "    with open(\"ch6_cell30_dev_feats2_test_raw.tsv\", \"w\") as feats:\n",
    "        feats.write(\"\\t\".join([\"name\"] + new_header + [\"population\"])+\"\\n\")\n",
    "        arr, target, names = xtest, ytest, test_names\n",
    "        for row in range(arr.shape[0]):\n",
    "            feats.write(\"\\t\".join(list(map(lambda v:str(v),\n",
    "                                       [names[row]]+list(arr[row,:])+[target[row]]))) +\"\\n\")\n",
    "\n",
    "# need to re-do tre after scaling\n",
    "xtrain = xtrain_non_tre\n",
    "xtest  = xtest_non_tre\n",
    "\n",
    "# SVRs need scaling\n",
    "xtrain_min = xtrain.min(axis=0); xtrain_max = xtrain.max(axis=0)\n",
    "# some can be zero if the column is constant in training\n",
    "xtrain_diff = xtrain_max - xtrain_min\n",
    "for idx in range(len(xtrain_diff)):\n",
    "    if xtrain_diff[idx] == 0.0:\n",
    "        xtrain_diff[idx] = 1.0\n",
    "xtrain_scaling = 1.0 / xtrain_diff\n",
    "xtrain -= xtrain_min; xtrain *= xtrain_scaling\n",
    "\n",
    "ytrain_min = ytrain.min(); ytrain_max = ytrain.max()\n",
    "ytrain_scaling = 1.0 / (ytrain_max - ytrain_min)\n",
    "ytrain -= ytrain_min; ytrain *= ytrain_scaling\n",
    "\n",
    "xtest -= xtrain_min; xtest *= xtrain_scaling\n",
    "ytest_orig = ytest.copy()\n",
    "ytest -= ytrain_min; ytest *= ytrain_scaling\n",
    "\n",
    "# redo TRE after scaling\n",
    "header, xtrain, ytrain, xtest, ytest, saved_full_TRE = doTRE(xtrain, ytrain, xtest, ytest)\n",
    "\n",
    "np.savez_compressed(\"ch6_cell30_scaled_full_TRE.npz\", \n",
    "                    saved_full_TRE = saved_full_TRE,\n",
    "                   xtrain_min      = xtrain_min,\n",
    "                   xtrain_scaling  = xtrain_scaling,\n",
    "                   ytrain_min      = ytrain_min,\n",
    "                   ytrain_scaling  = ytrain_scaling)\n",
    "with open(\"ch6_cell30_dev_feats2_train.tsv\", \"w\") as feats:\n",
    "    feats.write(\"\\t\".join([\"name\"] + header + [\"population\"])+\"\\n\")\n",
    "    arr, target, names = xtrain, ytrain, list(map(lambda t:t[2],train_data))\n",
    "    for row in range(arr.shape[0]):\n",
    "        feats.write(\"\\t\".join(list(map(lambda v:str(v),\n",
    "                                   [names[row]]+list(arr[row,:])+[target[row]]))) +\"\\n\")\n",
    "with open(\"ch6_cell30_dev_feats2_test.tsv\", \"w\") as feats:\n",
    "    feats.write(\"\\t\".join([\"name\"] + header + [\"population\"])+\"\\n\")\n",
    "    arr, target, names = xtest, ytest, test_names\n",
    "    for row in range(arr.shape[0]):\n",
    "        feats.write(\"\\t\".join(list(map(lambda v:str(v),\n",
    "                                   [names[row]]+list(arr[row,:])+[target[row]]))) +\"\\n\")\n",
    "\n",
    "\n",
    "# train\n",
    "print(\"Training on {:,} cities\".format(len(xtrain)))\n",
    "\n",
    "best_c = 100.0\n",
    "best_epsilon = 0.05\n",
    "PARAM_DO_GRID_SEARCH = False\n",
    "if PARAM_DO_GRID_SEARCH:\n",
    "    best_rmse = 1000\n",
    "    for c in [0.001, 0.01, 0.1, 1.0, 2.0, 5.0, 15.0, 100.0]:\n",
    "        svr_rbf = SVR(epsilon=0.01, C=c, gamma='auto')\n",
    "        svr_rbf.fit(xtrain, ytrain)\n",
    "        ytest_pred = svr_rbf.predict(xtest)\n",
    "        ytest_pred *= 1.0/ytrain_scaling\n",
    "        ytest_pred += ytrain_min\n",
    "        RMSE = math.sqrt(sum((ytest_orig - ytest_pred)**2) / len(ytest))\n",
    "        print(\"C\", c, \"RMSE\", RMSE)\n",
    "        if RMSE < best_rmse:\n",
    "            best_c = c\n",
    "            best_rmse = RMSE\n",
    "    print(\"Best C\", best_c,\"best RMSE\", best_rmse)        \n",
    "\n",
    "    best_epsilon = 0.1\n",
    "    best_rmse = 1000\n",
    "    for epsilon in [0.0001, 0.001, 0.01, 0.05, 0.1, 0.5, 0.0, 1.0, 10.0, 100.0]:\n",
    "        svr_rbf = SVR(epsilon=epsilon, C=best_c, gamma='auto')\n",
    "        svr_rbf.fit(xtrain, ytrain)\n",
    "        ytest_pred = svr_rbf.predict(xtest)\n",
    "        ytest_pred *= 1.0/ytrain_scaling\n",
    "        ytest_pred += ytrain_min\n",
    "        RMSE = math.sqrt(sum((ytest_orig - ytest_pred)**2) / len(ytest))\n",
    "        print(\"Epsilon\", epsilon, \"RMSE\", RMSE)\n",
    "        if RMSE < best_rmse:\n",
    "            best_epsilon = epsilon\n",
    "            best_rmse = RMSE\n",
    "    print(\"Best epsilon\", best_epsilon,\"best RMSE\", best_rmse)\n",
    "\n",
    "svr_rbf = SVR(epsilon=best_epsilon, C=best_c, gamma='auto')\n",
    "svr_rbf.fit(xtrain, ytrain)\n",
    "ytest_pred = svr_rbf.predict(xtest)\n",
    "ytest_pred *= 1.0/ytrain_scaling\n",
    "ytest_pred += ytrain_min\n",
    "RMSE = math.sqrt(sum((ytest_orig - ytest_pred)**2) / len(ytest))\n",
    "print(\"RMSE\", RMSE)\n",
    "\n",
    "print(\"Writing files for error analysis\")\n",
    "np.savez_compressed(\"ch6_cell30_feat2_svr.npz\", \n",
    "                    xtrain=xtrain, ytrain=ytrain, \n",
    "                    xtest=xtest, ytest=ytest,\n",
    "                    ytest_orig=ytest_orig,\n",
    "                    ytest_pred=ytest_pred,\n",
    "                    xtrain_scaling=xtrain_scaling,\n",
    "                    ytrain_scaling=ytrain_scaling)\n",
    "with open(\"ch6_cell30_test_names.tsv\", \"w\") as names:\n",
    "    for idx, name in enumerate(test_names):\n",
    "        names.write(\"{}\\t{}\\n\".format(idx, name))\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "plt.rcParams['figure.figsize'] = [20, 5]\n",
    "plt.plot(ytest_pred, label=\"predicted\", color='gray')\n",
    "plt.plot(ytest_orig, label=\"actual\", color='black')\n",
    "plt.ylabel('scaled log population')\n",
    "plt.savefig(\"ch6_cell30_svr_feat2.pdf\", bbox_inches='tight', dpi=300)\n",
    "plt.legend()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Adding back the original columns increases the number of columns to 577 from 314 but it reduces the RMSE to 0.3366 (otherwise it stays on 0.3514). This constitutes the best feature set to be used in the final evaluation.\n",
    "\n",
    "I'll now focus on selecting features that are stable under test error, that is, that can cope with a training set that is an incorrect sample of the feature.\n",
    "\n",
    "For that, I'll borrow a technique from privacy preserving data mining and introduce noise to the data by swapping the value of a given feature between two instances. Doing this swapping at a given percentage creates a test set that has been perturbed at a certain level. For these experiments, I will arbitrarily use a perturbation of 30%.\n",
    "\n",
    "For a conservative feature set, I will be willing to let go of features that, when perturbed at 30% produce a drop on performance greater than the drop resulting for keeping them.\n",
    "\n",
    "Sadly, as features are dependent, we'll need to retrain the base model every time a feature is deleted. Cell #31"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "2019-11-25T13:21:04.493754 === Total features 325 ===\n",
      "2019-11-25T13:21:26.182054 Base model trained\n",
      "2019-11-25T13:21:26.838139 Sampled model trained\n",
      "2019-11-25T13:21:26.838293 ablation,   0 out of 325\n",
      "2019-11-25T13:21:47.601105 ablation,  32 out of 325\n",
      "2019-11-25T13:22:08.500160 ablation,  64 out of 325\n",
      "2019-11-25T13:22:29.318456 ablation,  96 out of 325\n",
      "2019-11-25T13:22:50.153342 ablation, 128 out of 325\n",
      "2019-11-25T13:23:11.025659 ablation, 160 out of 325\n",
      "2019-11-25T13:23:31.847387 ablation, 192 out of 325\n",
      "2019-11-25T13:23:52.624032 ablation, 224 out of 325\n",
      "2019-11-25T13:24:13.417617 ablation, 256 out of 325\n",
      "2019-11-25T13:24:34.250685 ablation, 288 out of 325\n",
      "2019-11-25T13:24:55.108534 ablation, 320 out of 325\n",
      "2019-11-25T13:25:01.536526 perturbation at 0.3 finished for 3 samples\n",
      "2019-11-25T13:25:01.536699 perturbation,   0 out of 325\n",
      "2019-11-25T13:30:05.127399 perturbation,  32 out of 325\n",
      "2019-11-25T13:35:03.919973 perturbation,  64 out of 325\n",
      "2019-11-25T13:40:07.095331 perturbation,  96 out of 325\n",
      "2019-11-25T13:45:07.113808 perturbation, 128 out of 325\n",
      "2019-11-25T13:50:08.798181 perturbation, 160 out of 325\n",
      "2019-11-25T13:55:08.559696 perturbation, 192 out of 325\n",
      "2019-11-25T14:00:10.055907 perturbation, 224 out of 325\n",
      "2019-11-25T14:05:12.233867 perturbation, 256 out of 325\n",
      "2019-11-25T14:10:12.343686 perturbation, 288 out of 325\n",
      "2019-11-25T14:15:14.457111 perturbation, 320 out of 325\n",
      "2019-11-25T14:16:01.101299 Feature http://dbpedia.org/ontology/headquarter?inv#count ablated 0.0009274672328878353 perturbed 0.0028307071885098622 ratio 0.32764506221361295\n",
      "2019-11-25T14:16:01.121843 Feature http://dbpedia.org/ontology/timeZone#1@<http://dbpedia.org/resource/Philippine_Standard_Time> ablated 0.002247557793626953 perturbed 0.002763843223736906 ratio 0.8132001751489003\n",
      "2019-11-25T14:16:01.122189 Feature http://dbpedia.org/ontology/leaderTitle#1#TRE ablated 0.0001323788930047698 perturbed 0.000700552723678971 ratio 0.18896349772159696\n",
      "2019-11-25T14:16:01.122482 Feature http://dbpedia.org/ontology/type#1#TRE ablated 9.671300620706091e-05 perturbed 0.0005275024253355021 ratio 0.18334134889626244\n",
      "2019-11-25T14:16:01.122781 Feature http://dbpedia.org/ontology/utcOffset#2#TRE ablated 0.00010306182378183593 perturbed 0.0001874910024248777 ratio 0.5496894381538647\n",
      "Droppable features 5 out of 325\n",
      "2019-11-25T14:16:01.123193 Dropped http://dbpedia.org/ontology/type#1#TRE\n",
      "2019-11-25T14:16:01.174660 === Total features 324 ===\n",
      "2019-11-25T14:16:22.950512 Base model trained\n",
      "2019-11-25T14:16:23.619958 Sampled model trained\n",
      "2019-11-25T14:16:23.620119 ablation,   0 out of 324\n",
      "2019-11-25T14:16:44.949129 ablation,  32 out of 324\n",
      "2019-11-25T14:17:06.206845 ablation,  64 out of 324\n",
      "2019-11-25T14:17:27.473891 ablation,  96 out of 324\n",
      "2019-11-25T14:17:48.769531 ablation, 128 out of 324\n",
      "2019-11-25T14:18:10.095321 ablation, 160 out of 324\n",
      "2019-11-25T14:18:31.411787 ablation, 192 out of 324\n",
      "2019-11-25T14:18:52.719910 ablation, 224 out of 324\n",
      "2019-11-25T14:19:14.023926 ablation, 256 out of 324\n",
      "2019-11-25T14:19:35.313940 ablation, 288 out of 324\n",
      "2019-11-25T14:19:56.662283 ablation, 320 out of 324\n",
      "2019-11-25T14:20:02.512582 perturbation at 0.3 finished for 3 samples\n",
      "2019-11-25T14:20:02.512764 perturbation,   0 out of 324\n",
      "2019-11-25T14:25:02.365280 perturbation,  32 out of 324\n",
      "2019-11-25T14:30:00.944940 perturbation,  64 out of 324\n",
      "2019-11-25T14:34:57.904781 perturbation,  96 out of 324\n",
      "2019-11-25T14:39:57.777915 perturbation, 128 out of 324\n",
      "2019-11-25T14:44:57.481648 perturbation, 160 out of 324\n",
      "2019-11-25T14:49:55.597044 perturbation, 192 out of 324\n",
      "2019-11-25T14:54:53.113726 perturbation, 224 out of 324\n",
      "2019-11-25T14:59:52.254833 perturbation, 256 out of 324\n",
      "2019-11-25T15:04:52.506268 perturbation, 288 out of 324\n",
      "2019-11-25T15:09:52.198086 perturbation, 320 out of 324\n",
      "2019-11-25T15:10:29.797427 Feature http://dbpedia.org/ontology/country#1@<http://dbpedia.org/resource/Sweden> ablated 0.0006661093595425882 perturbed 0.0015691508509765306 ratio 0.4245030738300578\n",
      "2019-11-25T15:10:29.799311 Feature http://dbpedia.org/ontology/populationDensity#1 ablated 0.00010196205733148389 perturbed 0.00020525250642089432 ratio 0.4967640060014602\n",
      "Droppable features 2 out of 324\n",
      "2019-11-25T15:10:29.799785 Dropped http://dbpedia.org/ontology/country#1@<http://dbpedia.org/resource/Sweden>\n",
      "2019-11-25T15:10:29.808655 === Total features 323 ===\n",
      "2019-11-25T15:10:51.510013 Base model trained\n",
      "2019-11-25T15:10:52.159619 Sampled model trained\n",
      "2019-11-25T15:10:52.159778 ablation,   0 out of 323\n",
      "2019-11-25T15:11:12.904608 ablation,  32 out of 323\n",
      "2019-11-25T15:11:33.652429 ablation,  64 out of 323\n",
      "2019-11-25T15:11:54.393005 ablation,  96 out of 323\n",
      "2019-11-25T15:12:15.095731 ablation, 128 out of 323\n",
      "2019-11-25T15:12:35.869070 ablation, 160 out of 323\n",
      "2019-11-25T15:12:56.633469 ablation, 192 out of 323\n",
      "2019-11-25T15:13:17.466506 ablation, 224 out of 323\n",
      "2019-11-25T15:13:38.186955 ablation, 256 out of 323\n",
      "2019-11-25T15:13:58.962663 ablation, 288 out of 323\n",
      "2019-11-25T15:14:19.723642 ablation, 320 out of 323\n",
      "2019-11-25T15:14:24.805969 perturbation at 0.3 finished for 3 samples\n",
      "2019-11-25T15:14:24.806147 perturbation,   0 out of 323\n",
      "2019-11-25T15:19:21.228218 perturbation,  32 out of 323\n",
      "2019-11-25T15:24:18.040114 perturbation,  64 out of 323\n",
      "2019-11-25T15:29:16.511369 perturbation,  96 out of 323\n",
      "2019-11-25T15:34:13.985397 perturbation, 128 out of 323\n",
      "2019-11-25T15:39:12.060925 perturbation, 160 out of 323\n",
      "2019-11-25T15:44:08.263007 perturbation, 192 out of 323\n",
      "2019-11-25T15:49:05.444728 perturbation, 224 out of 323\n",
      "2019-11-25T15:54:01.894423 perturbation, 256 out of 323\n",
      "2019-11-25T15:59:01.978545 perturbation, 288 out of 323\n",
      "2019-11-25T16:04:03.259312 perturbation, 320 out of 323\n",
      "2019-11-25T16:04:31.048056 Feature http://dbpedia.org/ontology/areaLand#2 ablated 1.88743938718261e-05 perturbed 0.0010908843256056677 ratio 0.017301920495875568\n",
      "2019-11-25T16:04:31.048480 Feature http://dbpedia.org/ontology/elevation#1 ablated 0.0036923574838825246 perturbed 0.007026226673793268 ratio 0.5255107264976867\n",
      "2019-11-25T16:04:31.048738 Feature http://dbpedia.org/ontology/headquarter?inv#count ablated 0.000822406656177499 perturbed 0.0022617502509634235 ratio 0.3636151497395363\n",
      "2019-11-25T16:04:31.049028 Feature http://dbpedia.org/ontology/timeZone#1@<http://dbpedia.org/resource/Philippine_Standard_Time> ablated 0.0024296285228500025 perturbed 0.002935117625004313 ratio 0.827778928568981\n",
      "2019-11-25T16:04:31.049311 Feature http://dbpedia.org/ontology/leaderTitle#1#TRE ablated 0.000156618816564702 perturbed 0.0005469517852185746 ratio 0.2863484877412248\n",
      "Droppable features 5 out of 323\n",
      "2019-11-25T16:04:31.049675 Dropped http://dbpedia.org/ontology/areaLand#2\n",
      "2019-11-25T16:04:31.095231 === Total features 322 ===\n",
      "2019-11-25T16:04:52.794252 Base model trained\n",
      "2019-11-25T16:04:53.424141 Sampled model trained\n",
      "2019-11-25T16:04:53.424301 ablation,   0 out of 322\n",
      "2019-11-25T16:05:13.351756 ablation,  32 out of 322\n",
      "2019-11-25T16:05:33.308909 ablation,  64 out of 322\n",
      "2019-11-25T16:05:53.206302 ablation,  96 out of 322\n",
      "2019-11-25T16:06:13.194890 ablation, 128 out of 322\n",
      "2019-11-25T16:06:33.147183 ablation, 160 out of 322\n",
      "2019-11-25T16:06:53.098268 ablation, 192 out of 322\n",
      "2019-11-25T16:07:12.999528 ablation, 224 out of 322\n",
      "2019-11-25T16:07:32.922776 ablation, 256 out of 322\n",
      "2019-11-25T16:07:52.982776 ablation, 288 out of 322\n",
      "2019-11-25T16:08:13.029662 ablation, 320 out of 322\n",
      "2019-11-25T16:08:17.389080 perturbation at 0.3 finished for 3 samples\n",
      "2019-11-25T16:08:17.389263 perturbation,   0 out of 322\n",
      "2019-11-25T16:13:12.545119 perturbation,  32 out of 322\n",
      "2019-11-25T16:18:07.760685 perturbation,  64 out of 322\n",
      "2019-11-25T16:23:04.718894 perturbation,  96 out of 322\n",
      "2019-11-25T16:27:59.410018 perturbation, 128 out of 322\n",
      "2019-11-25T16:32:53.323151 perturbation, 160 out of 322\n",
      "2019-11-25T16:37:50.818356 perturbation, 192 out of 322\n",
      "2019-11-25T16:42:44.997881 perturbation, 224 out of 322\n",
      "2019-11-25T16:47:41.136197 perturbation, 256 out of 322\n",
      "2019-11-25T16:52:37.111684 perturbation, 288 out of 322\n",
      "2019-11-25T16:57:32.549776 perturbation, 320 out of 322\n",
      "2019-11-25T16:57:51.153114 Feature http://dbpedia.org/ontology/headquarter?inv#count ablated 0.0013303723614930263 perturbed 0.002678400924296831 ratio 0.4967039659465071\n",
      "2019-11-25T16:57:51.153547 Feature http://dbpedia.org/ontology/region#1@<http://dbpedia.org/resource/Midi-Pyrénées> ablated 0.00017285573278691402 perturbed 0.0002541867313277102 ratio 0.6800344450869853\n",
      "2019-11-25T16:57:51.153851 Feature http://dbpedia.org/ontology/timeZone#1@<http://dbpedia.org/resource/Philippine_Standard_Time> ablated 0.0037833149916854936 perturbed 0.004136512285338956 ratio 0.9146147117934836\n",
      "2019-11-25T16:57:51.154173 Feature http://dbpedia.org/ontology/populationAsOf#1#TRE ablated 5.619180415283864e-05 perturbed 9.206232238945233e-05 ratio 0.6103670067666748\n",
      "Droppable features 4 out of 322\n",
      "2019-11-25T16:57:51.154592 Dropped http://dbpedia.org/ontology/headquarter?inv#count\n",
      "2019-11-25T16:57:51.166271 === Total features 321 ===\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "2019-11-25T16:58:12.528099 Base model trained\n",
      "2019-11-25T16:58:13.181583 Sampled model trained\n",
      "2019-11-25T16:58:13.181736 ablation,   0 out of 321\n",
      "2019-11-25T16:58:33.904325 ablation,  32 out of 321\n",
      "2019-11-25T16:58:54.610138 ablation,  64 out of 321\n",
      "2019-11-25T16:59:15.285013 ablation,  96 out of 321\n",
      "2019-11-25T16:59:36.014337 ablation, 128 out of 321\n",
      "2019-11-25T16:59:56.781831 ablation, 160 out of 321\n",
      "2019-11-25T17:00:17.492602 ablation, 192 out of 321\n",
      "2019-11-25T17:00:38.177330 ablation, 224 out of 321\n",
      "2019-11-25T17:00:58.846890 ablation, 256 out of 321\n",
      "2019-11-25T17:01:19.613439 ablation, 288 out of 321\n",
      "2019-11-25T17:01:40.321650 ablation, 320 out of 321\n",
      "2019-11-25T17:01:44.149575 perturbation at 0.3 finished for 3 samples\n",
      "2019-11-25T17:01:44.149761 perturbation,   0 out of 321\n",
      "2019-11-25T17:06:38.764254 perturbation,  32 out of 321\n",
      "2019-11-25T17:11:32.003665 perturbation,  64 out of 321\n",
      "2019-11-25T17:16:26.601171 perturbation,  96 out of 321\n",
      "2019-11-25T17:21:22.294422 perturbation, 128 out of 321\n",
      "2019-11-25T17:26:17.450865 perturbation, 160 out of 321\n",
      "2019-11-25T17:31:12.911789 perturbation, 192 out of 321\n",
      "2019-11-25T17:36:08.240560 perturbation, 224 out of 321\n",
      "2019-11-25T17:41:02.736580 perturbation, 256 out of 321\n",
      "2019-11-25T17:45:57.929037 perturbation, 288 out of 321\n",
      "2019-11-25T17:50:50.702256 perturbation, 320 out of 321\n",
      "2019-11-25T17:50:59.788310 Feature http://dbpedia.org/ontology/department#1@<http://dbpedia.org/resource/Pas-de-Calais> ablated 1.3535612384707463e-05 perturbed 0.0013306511120747544 ratio 0.010172172301124601\n",
      "Droppable features 1 out of 321\n",
      "2019-11-25T17:50:59.790631 Dropped http://dbpedia.org/ontology/department#1@<http://dbpedia.org/resource/Pas-de-Calais>\n",
      "2019-11-25T17:50:59.842918 === Total features 320 ===\n",
      "2019-11-25T17:51:21.346834 Base model trained\n",
      "2019-11-25T17:51:21.999301 Sampled model trained\n",
      "2019-11-25T17:51:21.999452 ablation,   0 out of 320\n",
      "2019-11-25T17:51:42.694046 ablation,  32 out of 320\n",
      "2019-11-25T17:52:03.462559 ablation,  64 out of 320\n",
      "2019-11-25T17:52:24.109073 ablation,  96 out of 320\n",
      "2019-11-25T17:52:44.799092 ablation, 128 out of 320\n",
      "2019-11-25T17:53:05.610716 ablation, 160 out of 320\n",
      "2019-11-25T17:53:26.368710 ablation, 192 out of 320\n",
      "2019-11-25T17:53:47.018200 ablation, 224 out of 320\n",
      "2019-11-25T17:54:07.727404 ablation, 256 out of 320\n",
      "2019-11-25T17:54:28.491676 ablation, 288 out of 320\n",
      "2019-11-25T17:54:52.362356 perturbation at 0.3 finished for 3 samples\n",
      "2019-11-25T17:54:52.362539 perturbation,   0 out of 320\n",
      "2019-11-25T17:59:47.583158 perturbation,  32 out of 320\n",
      "2019-11-25T18:04:42.609576 perturbation,  64 out of 320\n",
      "2019-11-25T18:09:39.315901 perturbation,  96 out of 320\n",
      "2019-11-25T18:14:32.307300 perturbation, 128 out of 320\n",
      "2019-11-25T18:19:31.646431 perturbation, 160 out of 320\n",
      "2019-11-25T18:24:26.119408 perturbation, 192 out of 320\n",
      "2019-11-25T18:29:21.174792 perturbation, 224 out of 320\n",
      "2019-11-25T18:34:20.874156 perturbation, 256 out of 320\n",
      "2019-11-25T18:39:18.868541 perturbation, 288 out of 320\n",
      "2019-11-25T18:44:10.689198 Feature http://dbpedia.org/ontology/region#1@<http://dbpedia.org/resource/Nord-Pas-de-Calais> ablated 0.00026382005871017356 perturbed 0.0002930217859047346 ratio 0.9003428120390512\n",
      "2019-11-25T18:44:10.689593 Feature http://dbpedia.org/ontology/timeZone#1@<http://dbpedia.org/resource/Eastern_European_Time> ablated 0.0006883295292778868 perturbed 0.0009798518390997657 ratio 0.7024832753391435\n",
      "2019-11-25T18:44:10.689882 Feature http://dbpedia.org/ontology/timeZone#1@<http://dbpedia.org/resource/Philippine_Standard_Time> ablated 0.003919405571660484 perturbed 0.003944644014064913 ratio 0.9936018453593176\n",
      "Droppable features 3 out of 320\n",
      "2019-11-25T18:44:10.690323 Dropped http://dbpedia.org/ontology/timeZone#1@<http://dbpedia.org/resource/Eastern_European_Time>\n",
      "2019-11-25T18:44:10.700221 === Total features 319 ===\n",
      "2019-11-25T18:44:32.234996 Base model trained\n",
      "2019-11-25T18:44:32.885592 Sampled model trained\n",
      "2019-11-25T18:44:32.885743 ablation,   0 out of 319\n",
      "2019-11-25T18:44:52.911285 ablation,  31 out of 319\n",
      "2019-11-25T18:45:12.943576 ablation,  62 out of 319\n",
      "2019-11-25T18:45:32.936441 ablation,  93 out of 319\n",
      "2019-11-25T18:45:52.983884 ablation, 124 out of 319\n",
      "2019-11-25T18:46:13.007557 ablation, 155 out of 319\n",
      "2019-11-25T18:46:33.096562 ablation, 186 out of 319\n",
      "2019-11-25T18:46:53.156194 ablation, 217 out of 319\n",
      "2019-11-25T18:47:13.241975 ablation, 248 out of 319\n",
      "2019-11-25T18:47:33.254042 ablation, 279 out of 319\n",
      "2019-11-25T18:47:53.287071 ablation, 310 out of 319\n",
      "2019-11-25T18:48:02.171243 perturbation at 0.3 finished for 3 samples\n",
      "2019-11-25T18:48:02.171427 perturbation,   0 out of 319\n",
      "2019-11-25T18:52:44.687518 perturbation,  31 out of 319\n",
      "2019-11-25T18:57:28.070392 perturbation,  62 out of 319\n",
      "2019-11-25T19:02:10.266939 perturbation,  93 out of 319\n",
      "2019-11-25T19:06:53.403453 perturbation, 124 out of 319\n",
      "2019-11-25T19:11:48.346949 perturbation, 155 out of 319\n",
      "2019-11-25T19:16:31.979982 perturbation, 186 out of 319\n",
      "2019-11-25T19:21:15.780138 perturbation, 217 out of 319\n",
      "2019-11-25T19:25:57.895660 perturbation, 248 out of 319\n",
      "2019-11-25T19:30:41.987388 perturbation, 279 out of 319\n",
      "2019-11-25T19:35:23.337605 perturbation, 310 out of 319\n",
      "2019-11-25T19:36:44.859638 Feature http://dbpedia.org/ontology/department#1@<http://dbpedia.org/resource/Finistère> ablated 0.00013593849263049407 perturbed 0.00024706296418530727 ratio 0.5502180105332771\n",
      "2019-11-25T19:36:44.860015 Feature http://dbpedia.org/ontology/department#1@<http://dbpedia.org/resource/Haute-Garonne> ablated 0.0006367945464589297 perturbed 0.0020112021926608215 ratio 0.31662383264233135\n",
      "2019-11-25T19:36:44.860317 Feature http://dbpedia.org/ontology/populationAsOf#1@\"2013-12-31\"^^<http://www.w3.org/2001/XMLSchema#date> ablated 0.00031380754899235086 perturbed 0.0014953919016824635 ratio 0.2098497047090374\n",
      "2019-11-25T19:36:44.860680 Feature http://dbpedia.org/ontology/region#1#TRE ablated 9.056460346279493e-05 perturbed 0.00021524992415233868 ratio 0.4207416277587151\n",
      "2019-11-25T19:36:44.860971 Feature http://www.w3.org/2000/01/rdf-schema#seeAlso#3#TRE ablated 0.00020072408301086142 perturbed 0.0002105013432780467 ratio 0.9535525041554214\n",
      "Droppable features 5 out of 319\n",
      "2019-11-25T19:36:44.861382 Dropped http://dbpedia.org/ontology/populationAsOf#1@\"2013-12-31\"^^<http://www.w3.org/2001/XMLSchema#date>\n",
      "2019-11-25T19:36:45.019572 === Total features 318 ===\n",
      "2019-11-25T19:37:06.391448 Base model trained\n",
      "2019-11-25T19:37:07.045220 Sampled model trained\n",
      "2019-11-25T19:37:07.045374 ablation,   0 out of 318\n",
      "2019-11-25T19:37:27.131871 ablation,  31 out of 318\n",
      "2019-11-25T19:37:47.253811 ablation,  62 out of 318\n",
      "2019-11-25T19:38:07.306976 ablation,  93 out of 318\n",
      "2019-11-25T19:38:27.409822 ablation, 124 out of 318\n",
      "2019-11-25T19:38:47.533634 ablation, 155 out of 318\n",
      "2019-11-25T19:39:07.614966 ablation, 186 out of 318\n",
      "2019-11-25T19:39:27.686846 ablation, 217 out of 318\n",
      "2019-11-25T19:39:47.778393 ablation, 248 out of 318\n",
      "2019-11-25T19:40:07.864923 ablation, 279 out of 318\n",
      "2019-11-25T19:40:27.968146 ablation, 310 out of 318\n",
      "2019-11-25T19:40:36.236313 perturbation at 0.3 finished for 3 samples\n",
      "2019-11-25T19:40:36.236529 perturbation,   0 out of 318\n",
      "2019-11-25T19:45:17.523431 perturbation,  31 out of 318\n",
      "2019-11-25T19:49:57.402166 perturbation,  62 out of 318\n",
      "2019-11-25T19:54:50.171327 perturbation,  93 out of 318\n",
      "2019-11-25T19:59:33.063375 perturbation, 124 out of 318\n",
      "2019-11-25T20:04:13.985248 perturbation, 155 out of 318\n",
      "2019-11-25T20:08:53.768406 perturbation, 186 out of 318\n",
      "2019-11-25T20:13:33.388016 perturbation, 217 out of 318\n",
      "2019-11-25T20:18:13.589518 perturbation, 248 out of 318\n",
      "2019-11-25T20:22:53.090370 perturbation, 279 out of 318\n",
      "2019-11-25T20:27:34.706724 perturbation, 310 out of 318\n",
      "2019-11-25T20:28:46.383929 Feature http://dbpedia.org/ontology/department#1@<http://dbpedia.org/resource/Haute-Garonne> ablated 0.0005162292927192169 perturbed 0.0016265166156594901 ratio 0.3173833502524078\n",
      "Droppable features 1 out of 318\n",
      "2019-11-25T20:28:46.384510 Dropped http://dbpedia.org/ontology/department#1@<http://dbpedia.org/resource/Haute-Garonne>\n",
      "2019-11-25T20:28:46.394733 === Total features 317 ===\n",
      "2019-11-25T20:29:07.569353 Base model trained\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "2019-11-25T20:29:08.165877 Sampled model trained\n",
      "2019-11-25T20:29:08.166037 ablation,   0 out of 317\n",
      "2019-11-25T20:29:26.712355 ablation,  31 out of 317\n",
      "2019-11-25T20:29:45.119118 ablation,  62 out of 317\n",
      "2019-11-25T20:30:03.689109 ablation,  93 out of 317\n",
      "2019-11-25T20:30:22.226278 ablation, 124 out of 317\n",
      "2019-11-25T20:30:40.765717 ablation, 155 out of 317\n",
      "2019-11-25T20:30:59.321678 ablation, 186 out of 317\n",
      "2019-11-25T20:31:17.905135 ablation, 217 out of 317\n",
      "2019-11-25T20:31:36.504862 ablation, 248 out of 317\n",
      "2019-11-25T20:31:54.956919 ablation, 279 out of 317\n",
      "2019-11-25T20:32:13.476735 ablation, 310 out of 317\n",
      "2019-11-25T20:32:20.717909 perturbation at 0.3 finished for 3 samples\n",
      "2019-11-25T20:32:20.718091 perturbation,   0 out of 317\n",
      "2019-11-25T20:36:58.584525 perturbation,  31 out of 317\n",
      "2019-11-25T20:41:36.920274 perturbation,  62 out of 317\n",
      "2019-11-25T20:46:22.118936 perturbation,  93 out of 317\n",
      "2019-11-25T20:51:08.548078 perturbation, 124 out of 317\n",
      "2019-11-25T20:55:48.923326 perturbation, 155 out of 317\n",
      "2019-11-25T21:00:27.269818 perturbation, 186 out of 317\n",
      "2019-11-25T21:05:24.524497 perturbation, 217 out of 317\n",
      "2019-11-25T21:10:06.390253 perturbation, 248 out of 317\n",
      "2019-11-25T21:14:49.731397 perturbation, 279 out of 317\n",
      "2019-11-25T21:19:38.713480 perturbation, 310 out of 317\n",
      "2019-11-25T21:20:44.015651 Feature http://dbpedia.org/ontology/department#1@<http://dbpedia.org/resource/Finistère> ablated 3.886253194673728e-05 perturbed 0.0005379524209652153 ratio 0.07224157831097516\n",
      "2019-11-25T21:20:44.016535 Feature http://dbpedia.org/ontology/region#1@<http://dbpedia.org/resource/Haute-Normandie> ablated 3.2739117256319616e-05 perturbed 0.0005831928999149305 ratio 0.056137715773109075\n",
      "Droppable features 2 out of 317\n",
      "2019-11-25T21:20:44.017026 Dropped http://dbpedia.org/ontology/region#1@<http://dbpedia.org/resource/Haute-Normandie>\n",
      "2019-11-25T21:20:44.027917 === Total features 316 ===\n",
      "2019-11-25T21:21:05.505446 Base model trained\n",
      "2019-11-25T21:21:06.150688 Sampled model trained\n",
      "2019-11-25T21:21:06.150843 ablation,   0 out of 316\n",
      "2019-11-25T21:21:26.032294 ablation,  31 out of 316\n",
      "2019-11-25T21:21:45.939095 ablation,  62 out of 316\n",
      "2019-11-25T21:22:05.868594 ablation,  93 out of 316\n",
      "2019-11-25T21:22:25.739134 ablation, 124 out of 316\n",
      "2019-11-25T21:22:45.675613 ablation, 155 out of 316\n",
      "2019-11-25T21:23:05.660972 ablation, 186 out of 316\n",
      "2019-11-25T21:23:25.512030 ablation, 217 out of 316\n",
      "2019-11-25T21:23:45.310220 ablation, 248 out of 316\n",
      "2019-11-25T21:24:05.233605 ablation, 279 out of 316\n",
      "2019-11-25T21:24:25.161454 ablation, 310 out of 316\n",
      "2019-11-25T21:24:32.081030 perturbation at 0.3 finished for 3 samples\n",
      "2019-11-25T21:24:32.081213 perturbation,   0 out of 316\n",
      "2019-11-25T21:29:12.015567 perturbation,  31 out of 316\n",
      "2019-11-25T21:33:49.167385 perturbation,  62 out of 316\n",
      "2019-11-25T21:38:28.120821 perturbation,  93 out of 316\n",
      "2019-11-25T21:43:06.253766 perturbation, 124 out of 316\n",
      "2019-11-25T21:47:44.269183 perturbation, 155 out of 316\n",
      "2019-11-25T21:52:21.691520 perturbation, 186 out of 316\n",
      "2019-11-25T21:57:00.787898 perturbation, 217 out of 316\n",
      "2019-11-25T22:01:39.534595 perturbation, 248 out of 316\n",
      "2019-11-25T22:06:19.671841 perturbation, 279 out of 316\n",
      "2019-11-25T22:10:59.970293 perturbation, 310 out of 316\n",
      "Droppable features 0 out of 316\n",
      "2019-11-25T22:11:54.108416 Final number of features 316\n",
      "2019-11-25T22:11:54.108593 Start error 3.354327748503948\n",
      "2019-11-25T22:11:54.108753 Final error 3.3420276127247153\n"
     ]
    }
   ],
   "source": [
    "# CELL 31\n",
    "import random\n",
    "import datetime\n",
    "from sklearn.svm import SVR\n",
    "import numpy as np\n",
    "\n",
    "rand = random.Random(41) # 41 has areaLand and US; 39 areaLand and Brazil\n",
    "train_data = list()\n",
    "test_data  = list()\n",
    "header = None\n",
    "with open(\"ch6_cell30_dev_feats2_train.tsv\") as feats:\n",
    "    header = next(feats)\n",
    "    header = header.strip().split(\"\\t\")\n",
    "    header.pop(0) # name\n",
    "    header.pop()  # population\n",
    "    for line in feats:\n",
    "        fields = line.strip().split(\"\\t\")\n",
    "        logpop = float(fields[-1])\n",
    "        name  = fields[0]\n",
    "        feats = list(map(float,fields[1:-1]))\n",
    "        row   = (feats, logpop, name)\n",
    "        train_data.append(row)\n",
    "with open(\"ch6_cell30_dev_feats2_test.tsv\") as feats:\n",
    "    _ = next(feats)\n",
    "    for line in feats:\n",
    "        fields = line.strip().split(\"\\t\")\n",
    "        logpop = float(fields[-1])\n",
    "        name  = fields[0]\n",
    "        feats = list(map(float,fields[1:-1]))\n",
    "        row   = (feats, logpop, name)\n",
    "        test_data.append(row) \n",
    "               \n",
    "test_data  = sorted(test_data, key=lambda t:t[1])\n",
    "test_names = list(map(lambda t:t[2], test_data))\n",
    "\n",
    "xtrain = np.array(list(map(lambda t:t[0], train_data)))\n",
    "ytrain = np.array(list(map(lambda t:t[1], train_data)))\n",
    "xtest  = np.array(list(map(lambda t:t[0], test_data)))\n",
    "ytest  = np.array(list(map(lambda t:t[1], test_data)))\n",
    "\n",
    "# sample the data to finish in a reasonable amount of time\n",
    "selected = rand.sample(list(range(len(xtrain))), int(len(xtrain) * 0.2))\n",
    "xtrain = xtrain[selected,:]\n",
    "ytrain = ytrain[selected]\n",
    "\n",
    "selected = rand.sample(list(range(len(xtest))), int(len(xtest) * 0.2))\n",
    "xtest = xtest[selected,:]\n",
    "ytest = ytest[selected]\n",
    "\n",
    "dropped_features = list()\n",
    "rand = random.Random(42)\n",
    "\n",
    "best_c       = 100.0\n",
    "best_epsilon = 0.05\n",
    "PARAM_PERTURBATION_PERC = 0.3\n",
    "PARAM_SAMPLES           = 3\n",
    "PARAM_FIRST_SAMPLES     = 999\n",
    "start_error = None\n",
    "end_error   = None\n",
    "while True:\n",
    "    print(datetime.datetime.now().isoformat(), \"=== Total features\", xtrain.shape[1], \"===\")\n",
    "    # train base system\n",
    "    full_svr = SVR(epsilon=best_epsilon, C=best_c, gamma='auto')\n",
    "    full_svr.fit(xtrain, ytrain)\n",
    "    print(datetime.datetime.now().isoformat(), \"Base model trained\")\n",
    "    \n",
    "    # do wrapper\n",
    "    feature_loss = list()\n",
    "    selected     = rand.sample(list(range(len(xtrain))), int(len(xtrain) * 0.1))\n",
    "    xtrain_samp = xtrain[selected,:]\n",
    "    ytrain_samp = ytrain[selected]\n",
    "    \n",
    "    svr_rbf = SVR(epsilon=best_epsilon, C=best_c, gamma='auto')\n",
    "    svr_rbf.fit(xtrain_samp, ytrain_samp)\n",
    "    ytest_pred = svr_rbf.predict(xtest)\n",
    "    base_error = math.sqrt( sum((ytest_pred - ytest)**2) )\n",
    "    print(datetime.datetime.now().isoformat(), \"Sampled model trained\")\n",
    "    for feat in range(xtrain.shape[1]):\n",
    "        if feat % int(xtrain.shape[1] / 10) == 0:\n",
    "            print(datetime.datetime.now().isoformat(), \"ablation, {:>3} out of {:>3}\".format(feat, xtrain.shape[1]))\n",
    "        train_col = xtrain_samp[:,feat].copy()\n",
    "        xtrain_samp[:,feat] = 0.0 # nuke feature\n",
    "        svr_rbf = SVR(epsilon=best_epsilon, C=best_c, gamma='auto')\n",
    "        svr_rbf.fit(xtrain_samp, ytrain_samp)\n",
    "        ytest_pred = svr_rbf.predict(xtest)\n",
    "        error = math.sqrt( sum((ytest_pred - ytest)**2) )\n",
    "        feature_loss.append(base_error - error)\n",
    "        xtrain_samp[:,feat] = train_col # restore\n",
    "    while len(feature_loss) < xtrain.shape[1]:\n",
    "        feature_loss.append(999)\n",
    "\n",
    "    # do perturbation error\n",
    "    ytest_pred = full_svr.predict(xtest)\n",
    "    base_error = math.sqrt( sum( (ytest_pred - ytest)**2 ) )\n",
    "    if start_error is None:\n",
    "        start_error = base_error\n",
    "    end_error = base_error\n",
    "    \n",
    "    # perturbed samples\n",
    "    perturbed = list()\n",
    "    for _ in range(PARAM_SAMPLES):\n",
    "        xtest_perturbed = xtest.copy()\n",
    "        for _ in range(int(xtest.shape[0] * PARAM_PERTURBATION_PERC * 2)):\n",
    "            a = rand.randint(0, xtest.shape[0] - 1)\n",
    "            b = rand.randint(0, xtest.shape[0] - 1)\n",
    "            xtest_perturbed[ [a,b], :] = xtest_perturbed[ [b,a], :]\n",
    "        perturbed.append(xtest_perturbed)\n",
    "    print(datetime.datetime.now().isoformat(), \n",
    "          \"perturbation at {} finished for {} samples\".format(PARAM_PERTURBATION_PERC, PARAM_SAMPLES))\n",
    "\n",
    "    perturbed_error = list()\n",
    "    for feat in range(min(PARAM_FIRST_SAMPLES, xtrain.shape[1])):\n",
    "        if feat % int(xtrain.shape[1] / 10) == 0:\n",
    "            print(datetime.datetime.now().isoformat(), \"perturbation, {:>3} out of {:>3}\".format(feat, xtrain.shape[1]))\n",
    "        errors = list()\n",
    "        test_col = xtest[:,feat].copy()\n",
    "        for idx in range(PARAM_SAMPLES):\n",
    "            xtest[:,feat] = perturbed[idx][:,feat]\n",
    "            ytest_pred = full_svr.predict(xtest)\n",
    "            errors.append( base_error - math.sqrt( sum((ytest_pred - ytest)**2) ) )\n",
    "        xtest[:,feat] = test_col\n",
    "        perturbed_error.append(sorted(errors)[int(len(errors)/2)])\n",
    "    while len(perturbed_error) < xtrain.shape[1]:\n",
    "        perturbed_error.append(1)\n",
    "        \n",
    "    # pick the feature to drop\n",
    "    feat_to_drop = None\n",
    "    best_ratio = None\n",
    "    droppable_count = 0\n",
    "    for feat in range(min(PARAM_FIRST_SAMPLES, xtrain.shape[1])):\n",
    "        if perturbed_error[feat] > 0 and feature_loss[feat] > 0 and perturbed_error[feat] > feature_loss[feat]:\n",
    "            droppable_count += 1\n",
    "            ratio = feature_loss[feat] / perturbed_error[feat]\n",
    "            print(datetime.datetime.now().isoformat(), \"Feature\",header[feat], \"ablated\", feature_loss[feat], \n",
    "                  \"perturbed\",perturbed_error[feat], \"ratio\", ratio)\n",
    "            if feat_to_drop is None or best_ratio > ratio:\n",
    "                feat_to_drop = feat\n",
    "                best_ratio = ratio\n",
    "    print(\"Droppable features\", droppable_count, \"out of\", xtrain.shape[1])\n",
    "    if feat_to_drop is None:\n",
    "        break # done\n",
    "    \n",
    "    # drop feature\n",
    "    print(datetime.datetime.now().isoformat(), \"Dropped\", header[feat_to_drop])\n",
    "    dropped_features.append(header[feat_to_drop])\n",
    "    del header[feat_to_drop]\n",
    "    xtrain = np.delete(xtrain, feat_to_drop, 1)\n",
    "    xtest = np.delete(xtest, feat_to_drop, 1)\n",
    "    \n",
    "    if xtrain.shape[1] == 0:\n",
    "        break # shouldn't happen\n",
    "        \n",
    "print(datetime.datetime.now().isoformat(), \"Final number of features\",xtrain.shape[1])\n",
    "print(datetime.datetime.now().isoformat(), \"Start error\", start_error)\n",
    "print(datetime.datetime.now().isoformat(), \"Final error\", end_error)\n",
    "\n",
    "# write'em\n",
    "with open(\"ch6_cell31_features_dropped.txt\", \"w\") as dropped:\n",
    "    for featname in dropped_features:\n",
    "        dropped.write(featname + \"\\n\")\n",
    "\n",
    "with open(\"ch6_cell31_dev_train.tsv\", \"w\") as feats:\n",
    "    feats.write(\"\\t\".join([\"name\"] + header + [\"population\"])+\"\\n\")\n",
    "    arr, target, names = xtrain, ytrain, list(map(lambda t:t[2],train_data))\n",
    "    for row in range(arr.shape[0]):\n",
    "        feats.write(\"\\t\".join(list(map(lambda v:str(v),\n",
    "                                   [names[row]]+list(arr[row,:])+[target[row]]))) +\"\\n\")\n",
    "with open(\"ch6_cell31_dev_test.tsv\", \"w\") as feats:\n",
    "    feats.write(\"\\t\".join([\"name\"] + header + [\"population\"])+\"\\n\")\n",
    "    arr, target, names = xtest, ytest, test_names\n",
    "    for row in range(arr.shape[0]):\n",
    "        feats.write(\"\\t\".join(list(map(lambda v:str(v),\n",
    "                                   [names[row]]+list(arr[row,:])+[target[row]]))) +\"\\n\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The previous cell results suggest dropping countries but that is counter-intuitive as they are the most stable features. Based on this realization, I assemble a \"more conservative set\" using only counts and country features, with no TRE (Cell #32)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Read 3,424 countries\n",
      "Training on 35,940 cities\n",
      "RMSE 0.357821756992853\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x7f9310693750>"
      ]
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x360 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# CELL 32\n",
    "import bz2\n",
    "import re\n",
    "from sklearn.svm import SVR\n",
    "import numpy as np\n",
    "\n",
    "triple_re = re.compile(\"(<[^>]+>)\\s(<[^>]+>)\\s(.*) \\.\")\n",
    "\n",
    "def compressed_triples_gen(filename):\n",
    "    with bz2.BZ2File(filename,\"r\") as compressed:\n",
    "        for byteline in compressed:\n",
    "            if byteline[0] == ord('#'):\n",
    "                continue\n",
    "            line = byteline.decode(\"utf-8\")\n",
    "            s,v,o = triple_re.match(line).groups()\n",
    "            yield line,s,v,o\n",
    "            \n",
    "# determine country names, from DBpedia\n",
    "countries = list()\n",
    "for line, subj, verb, obj in compressed_triples_gen(\"instance_types_en.ttl.bz2\"):\n",
    "    if obj == '<http://dbpedia.org/ontology/Country>':\n",
    "        countries.append(subj[len(\"http://dbpedia.org/resource\"):])\n",
    "print(\"Read {:,} countries\".format(len(countries)))\n",
    "\n",
    "# read dataset\n",
    "train_data = list()\n",
    "test_data  = list()\n",
    "header = None\n",
    "devtrain = set()\n",
    "with open(\"devtrain.txt\") as train:\n",
    "    # for stable results due to a missing seed\n",
    "    for city in train:\n",
    "        devtrain.add(city.strip())\n",
    "with open(\"ch6_cell28_dev_feat1_filtered.tsv\") as feats:\n",
    "    header = next(feats)\n",
    "    header = header.strip().split(\"\\t\")\n",
    "    header.pop(0) # name\n",
    "    header.pop() # population\n",
    "    for line in feats:\n",
    "        fields = line.strip().split(\"\\t\")\n",
    "        logpop = float(fields[-1])\n",
    "        name = fields[0]\n",
    "        feats = list(map(float,fields[1:-1]))\n",
    "        row = (feats, logpop, name)\n",
    "        #if rand.random() < 0.2:\n",
    "        if name not in devtrain:\n",
    "            test_data.append(row) \n",
    "        else:\n",
    "            train_data.append(row)\n",
    "               \n",
    "train_names = list(map(lambda t:t[2], train_data))\n",
    "\n",
    "test_data  = sorted(test_data, key=lambda t:t[1])\n",
    "test_names = list(map(lambda t:t[2], test_data))\n",
    "\n",
    "xtrain = np.array(list(map(lambda t:t[0], train_data)))\n",
    "ytrain = np.array(list(map(lambda t:t[1], train_data)))\n",
    "xtest  = np.array(list(map(lambda t:t[0], test_data)))\n",
    "ytest  = np.array(list(map(lambda t:t[1], test_data)))\n",
    "\n",
    "# computed feature density area times\n",
    "approx_area_train = np.zeros( (xtrain.shape[0],) )\n",
    "approx_area_test  = np.zeros( (xtest.shape[0],)  )\n",
    "for idx, name in enumerate(header):\n",
    "    if \"#count\" not in name and (\"areaLand#\" in name or \"areaTotal#\" in name):\n",
    "        approx_area_train = np.max( [ approx_area_train, xtrain[:,idx] ], axis=0)\n",
    "        approx_area_test  = np.max( [ approx_area_test,  xtest[:,idx]  ], axis=0)\n",
    "density_idx = next(filter(lambda t:\"populationDensity#1\" in t[1], \n",
    "                     enumerate(header)))[0]\n",
    "computed_value_train = approx_area_train * xtrain[:,density_idx]\n",
    "computed_value_test  = approx_area_test  * xtest[:, density_idx]\n",
    "computed_defined_train = np.array( list(map(lambda x: 1.0 if x>0 else 0.0, \n",
    "                                            computed_value_train)) )\n",
    "computed_defined_test  = np.array( list(map(lambda x: 1.0 if x>0 else 0.0, \n",
    "                                            computed_value_test))  )\n",
    "computed_value_train = np.log(computed_value_train + 1)\n",
    "computed_value_test  = np.log(computed_value_test  + 1)\n",
    "\n",
    "# filter categories not containing a country\n",
    "to_delete = list()\n",
    "for idx, name in enumerate(header):\n",
    "    keep = True\n",
    "    if '@' in name:\n",
    "        keep = False\n",
    "        for country in countries:\n",
    "            if country in name:\n",
    "                keep = True\n",
    "                break\n",
    "    if not keep:\n",
    "        to_delete.append(idx)\n",
    "xtrain = np.delete(xtrain, to_delete, 1)\n",
    "xtest  = np.delete(xtest,  to_delete, 1)\n",
    "for idx in reversed(to_delete):\n",
    "    del header[idx]\n",
    "\n",
    "# add computed feature\n",
    "header.append(\"computed#defined\")\n",
    "header.append(\"computed#value\")\n",
    "xtrain = np.append(xtrain, computed_defined_train.reshape( (xtrain.shape[0],1) ), axis=1)\n",
    "xtrain = np.append(xtrain, computed_value_train.reshape(   (xtrain.shape[0],1) ), axis=1)\n",
    "xtest  = np.append(xtest,  computed_defined_test.reshape(  (xtest.shape[0], 1) ), axis=1)\n",
    "xtest  = np.append(xtest,  computed_value_test.reshape(    (xtest.shape[0], 1) ), axis=1)\n",
    "\n",
    "# write full file\n",
    "with open(\"ch6_cell32_dev_feat_conservative.tsv\",  \"w\") as f:\n",
    "    f.write(\"name\\t\" + \"\\t\".join(header) + '\\tpopulation\\n')\n",
    "    for row in range(xtrain.shape[0]):\n",
    "        f.write(train_names[row] + '\\t' + '\\t'.join(map(str,xtrain[row,:]))+ '\\t' + str(ytrain[row]) + '\\n')\n",
    "    for row in range(xtest.shape[0]):\n",
    "        f.write(test_names[row] +  '\\t' + '\\t'.join(map(str,xtest[row,:])) + '\\t' + str(ytest[row])  + '\\n')\n",
    "\n",
    "# train SVR\n",
    "xtrain_min = xtrain.min(axis=0); xtrain_max = xtrain.max(axis=0)\n",
    "# some can be zero if the column is constant in training\n",
    "xtrain_diff = xtrain_max - xtrain_min\n",
    "for idx in range(len(xtrain_diff)):\n",
    "    if xtrain_diff[idx] == 0.0:\n",
    "        xtrain_diff[idx] = 1.0\n",
    "xtrain_scaling = 1.0 / xtrain_diff\n",
    "xtrain -= xtrain_min; xtrain *= xtrain_scaling\n",
    "\n",
    "ytrain_min = ytrain.min(); ytrain_max = ytrain.max()\n",
    "ytrain_scaling = 1.0 / (ytrain_max - ytrain_min)\n",
    "ytrain -= ytrain_min; ytrain *= ytrain_scaling\n",
    "\n",
    "xtest -= xtrain_min; xtest *= xtrain_scaling\n",
    "ytest_orig = ytest.copy()\n",
    "ytest -= ytrain_min; ytest *= ytrain_scaling\n",
    "\n",
    "# train\n",
    "print(\"Training on {:,} cities\".format(len(xtrain)))\n",
    "\n",
    "best_c = 100.0\n",
    "best_epsilon = 0.05\n",
    "svr_rbf = SVR(epsilon=best_epsilon, C=best_c, gamma='auto')\n",
    "svr_rbf.fit(xtrain, ytrain)\n",
    "ytest_pred = svr_rbf.predict(xtest)\n",
    "ytest_pred *= 1.0/ytrain_scaling\n",
    "ytest_pred += ytrain_min\n",
    "RMSE = math.sqrt(sum((ytest_orig - ytest_pred)**2) / len(ytest))\n",
    "print(\"RMSE\", RMSE)\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "plt.rcParams['figure.figsize'] = [20, 5]\n",
    "plt.plot(ytest_pred, label=\"predicted\", color='gray')\n",
    "plt.plot(ytest_orig, label=\"actual\", color='black')\n",
    "plt.ylabel('scaled log population')\n",
    "plt.savefig(\"ch6_cell32_svr_feat_conservative.pdf\", bbox_inches='tight', dpi=300)\n",
    "plt.legend()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Follow ups left to the reader:\n",
    "\n",
    "* go back to the source data and encode the selected relations as sets\n",
    "* gps coordinates\n",
    "* add the countries explicitly\n",
    "* out-of-fold cluster the categories by their behaviour with the target class rather than by sheer counts\n",
    "* add an out-of-fold feature as the result of the decision tree (need to add \"ensemble systems as out-of-fold computable features\" to chapter 3).\n",
    "* final evaluation on held out"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Final systems:\n",
    "\n",
    "Assemble baseline system using feature stability and feat2 as good, with expected scores to get <- baseline: total count, plus area times density, country target rate encoded\n",
    "\n",
    "* number of relations, utc, density, areatotal, country with interpolated RFs (baseline)\n",
    "* baseline + more features including some target rate encoding, using a SVR (improved)\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "feateng",
   "language": "python",
   "name": "feateng"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}