DeepCTR

This project implements serval models of the papers on CTR prediction with easy-to-use call interfaces.

The goal is to make it possible for everyone to use complex models with model.fit()andmodel.predict().

Most of the models have been finished in keras. The tensorflow version will be added soon~ Please feel free to contact me if you have any questions!!

Support Model List

Model	Paper	Available Framework
AFM	[IJCAI 2017]Attentional Factorization Machines: Learning the Weight of Feature Interactions via Attention Networks	keras
DCN	[ADKDD 2017]Deep & Cross Network for Ad Click Predictions	keras,tensorflow
DeepFM	[IJCAI 2017]DeepFM: A Factorization-Machine based Neural Network for CTR Prediction	keras,tensorflow
MLR	[arxiv 2017]Learning Piece-wise Linear Models from Large Scale Data for Ad Click Prediction	keras,
NFM	[SIGIR 2017]Neural Factorization Machines for Sparse Predictive Analytics	keras

Keras Model

Requirements

• python3
• tensorflow==1.4.0
• keras==2.1.2

Quick Start

Source code keras_demo.py

data = pd.read_pickle("./demo/small_data.pkl")
sparse_features = [ "movie_id","user_id","gender","age","occupation","zip"]
target = ['rating']
#1.Label Encoding for sparse features,and Normalization for dense fetures
for feat in sparse_features:
  lbe = LabelEncoder()
  data[feat] = lbe.fit_transform(data[feat])
#2.count #unique features for each sparse field
sparse_feature_dim = {feat:data[feat].nunique() for feat in sparse_features}
#3.generate input data for model
model_input = [data[feat].values for feat in sparse_feature_dim] 
#4.Define Model,compile and train
model = NFM({"sparse":sparse_feature_dim,"dense":[]},final_activation='linear').model

model.compile("adam","mse",metrics=['mse'],)
history = model.fit(model_input,data[target],
        batch_size=256,epochs=5,verbose=2,validation_split=0.2,)

TensorFlow Model

Requirements

• python3
• tensorflow==1.4.0
• numpy==1.13.3
• scikit-learn==0.19.1

Design Notes

The base base class mimics the keras model to implement the following public methods, including:

• compile
• save_model
• load_mdel
• train_on_batch
• fit
• test_on_batch
• evaluate
• predict_on_batch
• predict

private methods:

• _create_optimizer
• _create_metrics
• _compute_sample_weight

At the same time, several abstract methods are designed:

• _get_input_data
• _get_input_target
• _get_output_target
• _get_optimizer_loss
• _build_graph

The subclass is required to call self._build_graph() at the end of the __init__ method to build the calculation graph.

TODO

• Add tf.summary.FileWriter
• Add custom metric function
• Add weighted loss function
• Encapsulate models with tf.estimator

Quick Start

Source code tf_demo.py

Experiment Result

see docs/README.md