This repository contains TensorFlow and pyTorch code and datasets for the paper:

Lianghao Xia, Chao Huang, Yong Xu, Jiashu Zhao, Dawei Yin, Jimmy Xiangji Huang (2022). Hypergraph Contrastive Collaborative Filtering, Paper in arXiv, Paper in ACM. In SIGIR'22, Madrid, Spain, July 11-15, 2022.

Hypergraph Contrastive Collaborative Filtering (HCCF) devises parameterized hypergraph neural network and hypergraph-graph contrastive learning, to relieve the over-smoothing issue for conventional graph neural networks, and address the sparse and skewed data distribution problem in collaborative filtering.

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

@inproceedings{hccf2022,
  author    = {Xia, Lianghao and
               Huang, Chao and
	       Xu, Yong and
	       Zhao, Jiashu and
	       Yin, Dawei and
	       Huang, Jimmy Xiangji},
  title     = {Hypergraph Contrastive Collaborative Filtering},
  booktitle = {Proceedings of the 45th International {ACM} {SIGIR} Conference on
               Research and Development in Information Retrieval, {SIGIR} 2022, Madrid,
               Spain, July 11-15, 2022.},
  year      = {2022},
}

The codes of HCCF are implemented and tested under the following development environment:

TensorFlow:

• python=3.6.12
• tensorflow=1.14.0
• numpy=1.16.0
• scipy=1.5.2

pyTorch:

• python=3.10.4
• torch=1.11.0
• numpy=1.22.3
• scipy=1.7.3

We utilized three datasets to evaluate HCCF: Yelp, MovieLens, and Amazon. Following the common settings of implicit feedback, if user $u_i$ has 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, validation set and testing set by 7:1:2.

Please unzip the datasets first. Also you need to create the History/ and the Models/ directories. The command to train HCCF on the Yelp/MovieLens/Amazon dataset is as follows. The commands specify the hyperparameter settings that generate the reported results in the paper.

• Yelp

python labcode_efficient.py --data yelp --temp 1 --ssl_reg 1e-4

• MovieLens

python labcode_efficient.py --data ml10m --temp 0.1 --ssl_reg 1e-6 --keepRate 1.0 --reg 1e-3

• Amazon

python labcode_efficient.py --data amazon --temp 0.1 --ssl_reg 1e-7 --reg 1e-2

Switch your working directory to torchVersion/, run python Main.py. The implementation has been improved in the torch code. You may need to adjust the hyperparameter settings. If you want to run HCCF on other datasets, we suggest you consider using a simplified version torchVersion/Model_sparse.py if your dataset is sparse. To do so, you should change the imported module in torchVersion/Main.py from Model to Model_sparse.

Important arguments:

• reg: It is the weight for weight-decay regularization. We tune this hyperparameter from the set {1e-2, 1e-3, 1e-4, 1e-5}.
• ssl_reg: This is the weight for the hypergraph-graph contrastive learning loss. The value is tuned from 1e-2 to 1e-8.
• temp: This is the temperature factor in the InfoNCE loss in our contrastive learning. The value is selected from {10, 3, 1, 0.3, 0.1}.
• keepRate: It denotes the rate to keep edges in the graph dropout, which is tuned from {0.25, 0.5, 0.75, 1.0}.
• leaky: The slope of the leakyReLU activation function. This parameter is recommended to tune from {1.0, 0.5, 0.1}.
• mult: A hyperparameter to manually tune the embedding magnitude of hypergraph NN in Model_sparse.py. Empirically, you can tune this parameter for the simplified version of HCCF from {1e-2, 1e-1, 1}.

This research is supported by the research grants from the Department of Computer Science & Musketeers Foundation Institute of Data Science at the University of Hong Kong, the Natural Sciences & Engineering Research Council (NSERC) of Canada.