hts-constrained-embeddings

This repository contains code to reproduce the experiments in "Forecasting Hierarchical Time Series with a Regularized Embedding Space" from the 6th Workshop on Mining and Learning from Time Series at KDD '20.

Data

The Australian travel flow data used in experiments can be downloaded here: https://robjhyndman.com/publications/hierarchical-tourism/.

Experiments

Run the experiment.py script with the following flags to reproduce experiments:

1. Initial Experiment:

   1. python experiment.py reproduce --metrics_file='data/metrics/baseline.txt' --output_path='data/preds_baseline' --serialize_path='data/models_baseline' --reconciled_path='data/reconciled_preds_baseline'

   2. in R: reconcile_all(in_dir = 'preds_baseline', out_dir = 'reconciled_preds_baseline')

   3. python experiment.py optimal_reconciliation --metrics_file='data/metrics/baseline.txt' --output_path='data/preds_baseline' --serialize_path='data/models_baseline' --reconciled_path='data/reconciled_preds_baseline'

2. Short Training Sequences:

   1. python experiment.py reproduce --train_size=54 --metrics_file='data/metrics/small.txt' --output_path='data/preds_small' --serialize_path='data/models_small' --reconciled_path='data/reconciled_preds_small'

   2. in R: reconcile_all(in_dir = 'preds_small', out_dir = 'reconciled_preds_small')

   3. python experiment.py optimal_reconciliation --train_size=54 --metrics_file='data/metrics/small.txt' --output_path='data/preds_small' --serialize_path='data/models_small' --reconciled_path='data/reconciled_preds_small'

3. Long Forecast Horizon:

   1. python experiment.py reproduce --horizon=24 --metrics_file='data/metrics/horizon.txt' --output_path='data/preds_horizon' --serialize_path='data/models_horizon' --reconciled_path='data/reconciled_preds_horizon'

   2. in R: reconcile_all(in_dir = 'preds_horizon', out_dir = 'reconciled_preds_horizon', horizon = 24)

   3. python experiment.py optimal_reconciliation --horizon=24 --metrics_file='data/metrics/horizon.txt' --output_path='data/preds_horizon' --serialize_path='data/models_horizon' --reconciled_path='data/reconciled_preds_horizon'

4. Relative Embedding Dimension: 0.5: python experiment.py reproduce --embed_dim_ratio=0.5 --metrics_file='data/metrics/edim05.txt' --output_path='data/preds_edim05' --serialize_path='data/models_edim05' --reconciled_path='data/reconciled_preds_edim05'

5. Relative Embedding Dimension: 2: python experiment.py reproduce --embed_dim_ratio=2 --metrics_file='data/metrics/edim2.txt' --output_path='data/preds_edim2' --serialize_path='data/models_edim2' --reconciled_path='data/reconciled_preds_edim2'

6. Relative Embedding Dimension: 4: python experiment.py reproduce --embed_dim_ratio=4 --metrics_file='data/metrics/edim4.txt' --output_path='data/preds_edim4' --serialize_path='data/models_edim4' --reconciled_path='data/reconciled_preds_edim4'

Plots

Run the plot.py script with the following flags to reproduce plots. Experiments must first be run and log data must be downloaded from tensorboard.

1. Initial Experiment: python plot.py --fig_title='baseline'

2. Short Training Sequences: python plot.py --fig_title='small'

3. Long Forecast Horizon: python plot.py --fig_title='horizon'