This is our Pytorch implementation for the paper:

Yunshan Ma, Yingzhi He, An Zhang, Xiang Wang, and Tat-Seng Chua (2022). CrossCBR: Cross-view Contrastive Learning for Bundle Recommendation, Paper in arXiv. In KDD'22, August 14–18, 2022, Washington, DC, USA.

Author: Yunshan Ma (yunshan.ma at u.nus.edu) and Yingzhi He (heyingzhi at u.nus.edu)

CrossCBR is a new recommendation model based on graph neural network and contrastive learning for bundle recommendation. By explicitly modeling the cooperative association between the item-view and bundle-view representations using an auxiliary contrastive loss, CrossCBR achieves great performance on three public bundle recommendation datasets: Youshu, NetEase, and iFashion.

If you want to use our codes and datasets in your research, please cite:

@inproceedings{CrossCBR2022,
  author    = {Yunshan Ma and
               Yingzhi He and
               An Zhang and
               Xiang Wang and
               Tat{-}Seng Chua},
  title     = {CrossCBR: Cross-view Contrastive Learning for Bundle Recommendation},
  booktitle = {{KDD}},
  pages     = {1233--1241},
  publisher = {{ACM}},
  year      = {2022}
}

• OS: Ubuntu 18.04 or higher version
• python == 3.7.3 or above
• supported(tested) CUDA versions: 10.2
• Pytorch == 1.9.0 or above

1. The entry script for training and evaluation is: train.py.
2. The config file is: config.yaml.
3. The script for data preprocess and dataloader: utility.py.
4. The model folder: ./models.
5. The experimental logs in tensorboard-format are saved in ./runs.
6. The experimental logs in txt-format are saved in ./log.
7. The best model and associate config file for each experimental setting is saved in ./checkpoints.

1. Decompress the dataset file into the current folder:

   tar -zxvf dataset.tgz

   Noted: for the iFashion dataset, we incorporate three additional files: user_id_map.json, item_id_map.json, and bundle_id_map.json, which record the id mappings between the original string-formatted id in the POG dataset and the integer-formatted id in our dataset. You may use the mappings to obtain the original content information of the items/outfits. We do not use any content information in this work.

2. Train CrossCBR on the dataset Youshu with GPU 0:

   python train.py -g 0 -m CrossCBR -d Youshu

   You can specify the gpu id and the used dataset by cmd line arguments, while you can tune the hyper-parameters by revising the configy file config.yaml. The detailed introduction of the hyper-parameters can be seen in the config file, and you are highly encouraged to read the paper to better understand the effects of some key hyper-parameters.