Code for Reinforced Neighborhood Selection Guided Multi-Relational Graph Neural Networks.

Hao Peng, Ruitong Zhang, Yingtong Dou, Renyu Yang, Jingyi Zhang, Philip S. Yu.

The repository is organized as follows:

• data/: dataset folder
  • YelpChi.zip: Data of the dataset Yelp;
  • Amazon.zip: Data of the dataset Amazon;
  • Mimic.zip: Data of the dataset Mimic;
• log/: log folder
• model/: model folder
  • graphsage.py: model code for vanilla GraphSAGE model;
  • layers.py: RioGNN layers implementations;
  • model.py: RioGNN model implementations;
• RL/: RL folder
  • actor_critic.py: RL algorithm, Actor-Critic;
  • rl_model.py: RioGNN RL Forest implementations;
• utils/: functions folder
  • data_process.py: transfer sparse matrix to adjacency lists;
  • utils.py: utility functions for data i/o and model evaluation;
• train.py: training and testing all models

We build different multi-relational graphs for experiments in two task scenarios and three datasets:

To run RioGNN on your datasets, you need to prepare the following data:

• Multiple-single relation graphs with the same nodes where each graph is stored in scipy.sparse matrix format, you can use sparse_to_adjlist() in utils.py to transfer the sparse matrix into adjacency lists used by RioGNN;
• A numpy array with node labels. Currently, RioGNN only supports binary classification;
• A node feature matrix stored in scipy.sparse matrix format.

You can download the project and and run the program as follows:

Google Drive: https://drive.google.com/file/d/1WvYtNSHcvSQr8fzI9ykpgjMBSPwCTW0h/view?usp=sharing

Baidu Cloud: https://pan.baidu.com/s/1iyaOqnkyYGqo1Mdwt4QYnQ Password: vbwn

* Note that all datasets need to be unzipped in the folder \data first;

pip3 install -r requirements.txt

python data_process.py

python train.py

* To run the code, you need to have at least Python 3.6 or later versions.

• Our model supports both CPU and GPU mode, you can change it through parameter --use_cuda and --device:
• Set the --data as yelp, amazon or mimic to change different dataset.
• Parameter --num_epochs is used to set the maximum number of iterative epochs. Note that the model will stop early when reinforcement learning has explored all depths.
• The default value of parameter --ALAPHA is 10, which means that the accuracy of different depths of reinforcement learning tree will be progressive with 0.1, 0.01, 0.001, etc. If you want to conduct more width and depth experiments, please adjust here.

* For other dataset and parameter settings, please refer to the arg parser in train.py.

Our preliminary work, CAmouflage-REsistant Graph Neural Network (CARE-GNN), is a GNN-based fraud detector based on a multi-relation graph equipped with three modules that enhance its performance against camouflaged fraudsters.

If you use our code, please cite the paper below:

@article{peng2021reinforced,
  title={Reinforced Neighborhood Selection Guided Multi-Relational Graph Neural Networks},
  author={Peng, Hao and Zhang, Ruitong and Dou, Yingtong and Yang, Renyu and Zhang, Jingyi and Yu, Philip S.},
  journal={ACM Transactions on Information Systems (TOIS)},
  year={2021}
}