THACIL Network

This repo is our implementation for paper: Temporal Hierarchical Attention at Category- and Item-Level for Micro-Video Click-Through Prediction

Please cite our MM'18 paper if you use our codes. Thanks!

@inproceedings{chen2018thacil,
  title={Temporal Hierarchical Attention at Category- and Item-Level for Micro-Video Click-Through Prediction},
  author={Chen, Xusong and Liu, Dong and Zha, Zheng-Jun and Zhou, Wengang and Xiong, Zhiwei and Li, Yan},
  booktitle={MM},
  year={2018}
}

Environment Setting

• Python 3.6 (Anaconda3)
• Tensorflow 1.4

Usage

Download dataset and preprocess

• Step 1: Download our dataset, named MicroVideo-1.7M (password: ms7x), which has 12,737,619 interactions that 10,986 users have made on 1,704,880 micro-videos. And move it to data/input/ folder.
• Step 2: Generate intermediate file for training, and move it to data/input/ folder.
  • Generate user_click_ids.npy:

    python generate_data.py --train-data-path ../../data/input/train_data.csv --save-path ../../data/input/
    

Training and Evaluation

• Training

cd src
python launcher.py  --phase train

• Evaluation

cd src
python launcher.py  --phase test

Acknowledgements

Some implementations consult the ATRank, and DiSAN.