DVR: Micro-Video Recommendation Optimizing Watch-Time-Gain under Duration Bias

This is the official implementation of our MM'22 paper:

Yu Zheng, Chen Gao, Jingtao Ding, Lingling Yi, Depeng Jin, Yong Li, Meng Wang, DVR: Micro-Video Recommendation Optimizing Watch-Time-Gain under Duration Bias, In Proceedings of the ACM Multimedia 2022.

The code is tested under a Linux desktop with TensorFlow 2.3.0 and Python 3.7.9.

Please cite our paper if you use this repository.

@inproceedings{zheng2022dvr,
  title={DVR: Micro-Video Recommendation Optimizing Watch-Time-Gain under Duration Bias},
  author={Zheng, Yu and Gao, Chen and Ding, Jingtao and Yi, Lingling and Jin, Depeng and Li, Yong and Wang, Meng},
  booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
  pages={334--345},
  year={2022}
}

Datasets

Unzip the compressed data files in examples/data/ with the following commands:

cd ./examples/data/
unzip kuaishou_video_debias.csv.zip
unzip wechat_video_debias.csv.zip

Model Training

Use the following commands for model training.

To train a basic DeepFM model on Kuaishou dataset:

cd ./examples
python run_video_debias.py --name kuaishou-deepfm --model DeepFM --dataset kuaishou 

To train a DVR- version of DeepFM model on Kuaishou dataset:

cd ./examples
python run_video_debias.py --name kuaishou-deepfm-dvrminus --model DeepFM --dataset kuaishou --post_transform

To train a DVR version of DeepFM model on Kuaishou dataset:

cd ./examples
python run_video_debias.py --name kuaishou-deepfm-dvr --model DeepFM --dataset kuaishou --train_target gain --remove_duration_feature --disentangle --disentangle_loss_weight 0.1

You can check the FLAGS in examples/run_video_debias.py to explore other experimental settings.

Note

The implemention is based on DeepCTR.