purpose

learning, try to reproduce the results of the paper

environment

• python 3.6
• tensorflow 2.0(keras API)
• tqdm

recsys algorithms

matrix factorization

reference algorithm

matrix factorization with temporal dynamics and bias(MF_Alg: Matrix Factorization Techniques for Recommender Systems)

training script

• training dataset

netflix dataset

• dataset processing

utils/netflix.py

two kinds of dataset api: tfrecord and tensorarray; big dataframe multiprocessing process

• train script

train/run_basic_exp.py

evaluation result

MF_Alg: 49ms/step - loss: 0.8175 - mean_squared_error: 0.7460 - val_loss: 1.0178 - val_mean_squared_error: 0.9235

sequence recommendation system(for learning purpose)

algorithms

DNN, LSTM, LSTM with attention

training script

• dataset

movielens 20M

sequence processing: utils/seq_process.py

• train script: train/run_seq_exp.py

• custom keras metrics

NCG MRR

• evaluation

dnn loss: tf.Tensor(0.26367342, shape=(), dtype=float32) ---epoch: 0, mrr:0.0223

LSTM: loss: tf.Tensor(0.13960448, shape=(), dtype=float32) ---epoch: 3, mrr:0.0433

attention: test OOM

reference

spotlight https://github.com/maciejkula/spotlight

excellent evaluation script: https://github.com/microsoft/recommenders/tree/9b6c065a77c07a60947e894a6b516be5cef14ca1/reco_utils/evaluation

GRU4Rec

evaluation

• Recall@20 epoch 5: 0.620944
• MRR@20 epoch 5: 0.279456

reference:

https://github.com/pcerdam/KerasGRU4Rec

my works

data proessing masks bug fix

working with tensorflow 2.0

CTR Models

factorization machine

• dataset

criteo: 45000000

• data processing

prepare_criteo_data.py: prepare the dataset

criteo_data.py: two kinds of tensorflow dataset api, textlinedataset and tfrecords

• evaluation

20429/20429 [==============================] - 3155s 154ms/step - loss: 0.5145 - auc_t: 0.7551 - val_loss: 0.5160 - val_auc_t: 0.7530

deepfm

*evaluation

163439/163439 [==============================] - 11687s 72ms/step - loss: 0.4795 - auc_t: 0.7867 - val_loss: 0.4795 - val_auc_t: 0.7879

TODO

optimizing the fm and deepfm, DIN network