This repository is an official PyTorch(Geometric) implementation of CAGCN (also includes the baseline implementation of MF/NGCF/LightGCN) in "Collaboration-Aware Graph Convolutional Networks for Recommendation Systems".

If you use this code, please consider citing:

@inproceedings{CAGCN,
  author={Wang, Yu and Zhao, Yuying and Zhang, Yi and Derr, Tyler},
  title={Collaboration-Aware Graph Convolutional Networks for Recommendation Systems},
  booktitle={Arxiv},
  year = {2022}
}

Most GNN-based recommendation systems perform message-passing by directly applying traditional GNN-convolutions. How does the message-passing captures collaborative effect? Are the message-passing by traditional GNN-convolutions really beneficial to users' ranking? This paper will take you to demystify the collaborations captured by traditional GNN-convolutions.

Specifically, we demystify the collaborations captured by LightGCN by answering the following two questions:

• How does the message-passing captures collaborative effect? We find that the L-layer LightGCN-based message-passing captures and leverages collaborations between nodes within L-hops neighborhoods of the center user and item. We also theoretically derive the strength of the captured collaborations.

• Does the captured collaboration really help the prediction of users' ranking? We propose a new recommendation-tailored topological metric, Common Interacted Ratio (CIR), and empirically find that higher CIR leads to more benefits to users' ranking.

Based on our theoretical and empirical analysis, we incorporate CIR into the message-passing and ultimately propose a novel class of Collaboration-Aware Graph Convolutional Networks, namely Collaboration-Aware Graph Convolutional Network (CAGCN) and its augmented version (CAGCN*), both of which are able to selectively pass information of neighbors based on their CIR via the designed Collaboration-Aware Graph Convolution. We also propose a brandly new type of graph isomorphism for bipartite graphs and theoretically prove that the designed CAGCN* can go beyond 1-WL test in distinguishing subtree-isomorphic(subgraph-isomorphic) graphs yet not bipartite-subgraph-isomorphic graphs. The whole framework and the superiority of CAGCN* over 1-WL are as follows:

The default version of python we use is 3.8.10. Please install all necessary python packages via:

- Pytorch 1.11.0 with Cuda 11.3
- Pytorch-geometric 2.0.4
- Torch-scatter 2.0.9
- Prettytable 3.2.0

We demonstrate the superority of CAGCN(*) on six datasets: Gowalla, Yelp2018, Amazon-book, Ml-1M, Loseit and Worldnews.

• The train1.txt/test1.txt: the observed and unobserved user-item interaction pairs.
• co_ratio_edge_weight_x.pt: the precalculated Common Interacted Ratio (CIR) based on x, which is selected from: Jaccard Similarity (JC), Salton Cosine Similarity (SC), Leicht-Holme-Nerman (LHN), and Common Neighbors (CN).
• The correct directory structure should be set as follows:

├── data
│   ├── amazon
│   │   ├── co_ratio_edge_weight_co.pt
│   │   ├── co_ratio_edge_weight_jc.pt
│   │   ├── co_ratio_edge_weight_lhn.pt
│   │   ├── co_ratio_edge_weight_sc.pt
│   │   ├── test1.txt
│   │   └── train1.txt
│   ├── gowalla
│   │   ├── co_ratio_edge_weight_co.pt
│   │   ├── co_ratio_edge_weight_jc.pt
│   │   ├── co_ratio_edge_weight_lhn.pt
│   │   ├── co_ratio_edge_weight_sc.pt
│   │   ├── test1.txt
│   │   └── train1.txt
│   ├── loseit
│   │   ├── co_ratio_edge_weight_co.pt
│   │   ├── co_ratio_edge_weight_jc.pt
│   │   ├── co_ratio_edge_weight_lhn.pt
│   │   ├── co_ratio_edge_weight_sc.pt
│   │   ├── test1.txt
│   │   └── train1.txt
│   ├── ml-1m
│   │   ├── co_ratio_edge_weight_co.pt
│   │   ├── co_ratio_edge_weight_jc.pt
│   │   ├── co_ratio_edge_weight_lhn.pt
│   │   ├── co_ratio_edge_weight_sc.pt
│   │   ├── test1.txt
│   │   └── train1.txt
│   ├── worldnews
│   │   ├── co_ratio_edge_weight_co.pt
│   │   ├── co_ratio_edge_weight_jc.pt
│   │   ├── co_ratio_edge_weight_lhn.pt
│   │   ├── co_ratio_edge_weight_sc.pt
│   │   ├── test1.txt
│   │   └── train1.txt
│   └── yelp2018
│       ├── co_ratio_edge_weight_co.pt
│       ├── co_ratio_edge_weight_jc.pt
│       ├── co_ratio_edge_weight_lhn.pt
│       ├── co_ratio_edge_weight_sc.pt
│       ├── test1.txt
│       └── train1.txt
├── dataprocess.py
├── evaluation.py
├── main_fusion.py
├── main.py
├── model.py
├── parse.py
├── run_amazon.sh
├── run_gowalla.sh
├── run_loseit.sh
├── run_ml1m.sh
├── run_world_news.sh
├── run_yelp.sh
└── utils.py

Note that we also provide the code for readers to pre-calculate the co_ratio_edge_weight_x.pt of their own datasets. To use it, please set up a separate dataset repo with the training and testing interactions and create your own bash.sh file:

bash run_xxx.sh

We have matrix-formed and node-wise calculation of CIR. The matrix-formed way is faster while requires more RAM while the node-wise calculation is slower but requires less RAM. To vary between these two modes, change the line in main.py/main_fusion.py as follows:

if args.dataset in ['amazon']:
    cal_trend = co_ratio_deg_user_jacard_sp
else:
    cal_trend = co_ratio_deg_user_jacard

• co_ratio_deg_user_jacard_sp: node-wise calculation
• co_ratio_deg_user_jacard: matrix-formed calculation

Here we list the performance of our models CAGCN(*) with different topological variants. To reproduce the performance and running time in the following Table, please run the following commands:

bash run_gowalla.sh
bash run_yelp.sh
bash run_amazon.sh
bash run_ml1m.sh
bash run_loseit.sh
bash run_worldnews.sh

[1] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, Meng Wang. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. SIGIR 2020.
[2] Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, Tat-Seng Chua. Neural graph collaborative filtering. SIGIR 2019.
[3] Tinglin Huang, Yuxiao Dong, Ming Ding, Zhen Yang, Wenzheng Feng, Xinyu Wang, Jie Tang. MixGCF: An Improved Training Method for Graph Neural Network-based Recommender Systems. KDD 2021.