This repository provides a reference implementation of MHGCN as described in the paper:
Multiplex Heterogeneous Graph Convolutional Network
Pengyang Yu, Chaofan Fu, Yanwei Yu, Chao Huang, Zhongying Zhao, Junyu Dong.
KDD'22
Available at https://doi.org/10.1145/3534678.3539482
Recent versions of the following packages for Python 3 are required:
- numpy==1.21.2
- torch==1.9.1
- scipy==1.7.1
- scikit-learn==0.24.2
- pandas==0.25.0
The used datasets are available at:
- Alibaba https://tianchi.aliyun.com/competition/entrance/231719/information/
- Amazon http://jmcauley.ucsd.edu/data/amazon/
- Aminer https://github.com/librahu/HIN-Datasets-for-Recommendation-and-Network-Embedding/tree/master/Aminer
- IMDB https://github.com/seongjunyun/Graph_Transformer_Networks
- DBLP https://www.dropbox.com/s/yh4grpeks87ugr2/DBLP_processed.zip?dl=0
We compress the data set into a mat format file, which includes the following contents.
- edges: array of subnetworks after coupling, each element in the array is a subnetwork.
- features: attributes of each node in the network.
- labels: label of labeled points.
- train: index of training set points for node classification.
- valid: index of validation set points for node classification.
- test: index of test set points for node classification.
In addition, we also sample the positive and negative edges in the network, and divide them into three text files: train, valid and test for link prediction.
First, you need to determine the data set. If you want to do node classification tasks, you need to modify the data set path in Node_classification.py. If you want to do link prediction, you need to modify the dataset path in Link_prediction.py.
Second, you need to modify the number of weights in Model.py. The number of weights should be the number of sub-networks after decoupling.
Finally, you need to determine the sub-network and the number of sub-networks in Decoupling_matrix_aggregation.py.
Execute the following command to run the node classification task:
python Node_Classification.py
Execute the following command to run the link prediction task:
python Link_Prediction.py
If you find MHGCN useful in your research, please cite the following paper:
@inproceedings{yu2022multiplex,
title={Multiplex Heterogeneous Graph Convolutional Network},
author={Yu, Pengyang and Fu, Chaofan and Yu, Yanwei and Huang, Chao and Zhao, Zhongying and Dong, Junyu},
booktitle={Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
pages={2377--2387},
year={2022}
}