Torch version is available now! This repository contains pyTorch code and datasets for the paper: Heterogeneous Graph Contrastive Learning for Recommendation, *Paper in arXiv, Paper in ACM. In WSDM'23, Singapore, 2023
Heterogeneous Graph Contrastive Learning for Recommendation (HGCL) advances the recommender system with heterogeneous graph contrastive learning. HGCL integrates meta network with contrastive learning for adaptive augmentation to enable user-specific and item-specific knowledge transfer. It advances graph contrastive learning with customized cross-view augmentation.
The codes of HGCL are implemented and tested under the following development environment: pyTorch: Python=3.7.10 Torch=1.8.1 Numpy=1.20.3 Scipy=1.6.2
We utilized three datasets to evaluate HGCL: Yelp, Epinions, and CiaoDVD. Following the common settings of implicit feedback, if user u_ihas rated item v_j, then the element (u_i,v_j) is set as 1, otherwise 0. We filtered out users and items with too few interactions. The datasets are divided into training set and testing set by 1: (n-1).
You can download all three datasets from Google Drive. Feel free to fire an issue if this link doesn't work.
Please unzip the datasets first. Also you need to create the History/ and the Models/ directories. The command to train HGCL on the Yelp/Epinions/CiaoDVD dataset is as follows. The commands specify the hyperparameter settings that generate the reported results in the paper.
- Yelp
python main.py --dataset Yelp --ssl_temp 0.5 --ssl_ureg 0.06 --ssl_ireg 0.07 --lr 0.058 --reg 0.05 --ssl_beta 0.45 --rank 3
- Epinions
python main.py --dataset Epinions --ssl_temp 0.5 --ssl_ureg 0.04 --ssl_ireg 0.05 --lr 0.055 --reg 0.043 --ssl_beta 0.32 --rank 3
- CiaoDVD
python main.py --dataset CiaoDVD --ssl_temp 0.6 --ssl_ureg 0.04 --ssl_ireg 0.05 --lr 0.055 --reg 0.065 --ssl_beta 0.3 --rank 3
--ssl_tempIt is the temperature factor in the InfoNCE loss in our contrastive learning. The value is selected from {0.1, 0.3, 0.45, 0.5, 0.55,0.6, 0.65}.--ssl_ureg, ssl_iregThey are the weights for the contrastive learning loss of user’s and item’s aspect respectively. The value of this pair are tuned from {(3e-2,4e-2),( 4e-2,5e-2),( 5e-2,6e-2), (6e-2,7e-2),( 7e-2,8e-2)}.--lrThe learning rate of the mode. We tuned it from {1e-2, 3e-2, 4e-2, 4.5e-2, 5e-2, 5.5e-2, 6e-2}.--RegIt is the weight for weight-decay regularization. We tune this hyperparameter from the set {1e-2, 3e-2, 4.3e-2, 5e-2, 6e-2, 6.5e-2, 6.8e-2}.--ssl_betaThis is the balance cofficient of the total contrastive loss , which is tuned from{0.2, 0.27, 0.3, 0.32, 0.4, 0.45, 0.48, 0.5}.--rankA hyperparameter of the dimension of low rank matrix decomposition, This parameter is recommended to tune from{1, 2, 3, 4, 5}.
Performance comparison of all methods on different datasets in terms of NDCG and HR:
If you find this work helpful to your research, please kindly consider citing our paper.
@inproceedings{chen2023heterogeneous,
title={Heterogeneous graph contrastive learning for recommendation},
author={Chen, Mengru and Huang, Chao and Xia, Lianghao and Wei, Wei and Xu, Yong and Luo, Ronghua},
booktitle={Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining},
pages={544--552},
year={2023}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent