MBGCN

This is our implementation of paper:

Bowen Jin, Chen Gao, Xiangnan He, Depeng Jin and Yong Li. 2020. Multi-behavior Recommendation with Graph Convolutional Networks. In SIGIR'20.

Please cite our SIGIR'20 paper if you use our codes. Thanks!

@inproceedings{jin2020multi,
  title={Multi-behavior Recommendation with Graph Convolution Networks},
  author={Jin, Bowen and Gao, Chen and He, Xiangnan and Jin, Depeng and Li, Yong},
  booktitle={43nd International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year={2020},
}

Author: Bowen Jin ([email protected])

Enviroments

• python
• pytorch
• numpy
• visdom

Sampling

Construct positive and negative item pair for BPR loss by running:

cd Tmall
mkdir sample_file
cd ..
python sample.py --path Tmall

Running

Visdom

Open a visdom port by running

visdom -port 33337

Make port forwarding and then you can visit localhost:33337 with explorer.

Pretrain

Train MF first by running:

bash MF.sh

Train

Change 'pretrain_path' in MBGCN.sh to the path where the best MF model located.

Train MBGCN by running:

bash MBGCN.sh

Note

We change sampling method from sampling online using DataLoader with 8 workers to sampling offline and save the pairs in .txt in advance. As a result, with code here, all BPR-based method including our MBGCN will get better performance compared with the performance in our paper.