Rayleigh Quotient Graph Neural Networks for Graph-level Anomaly Detection

Tested Environment

• Ubuntu 20.04
• Python 3.7.15
• Sklearn 1.0.2
• Pytorch 1.12.1
• Numpy 1.21.6
• Torch_geometric 2.2.0
• Scipy 1.7.3

Datasets

Download zip files from TUDatasets and unzip them in datasets/.

Directory Structure

├── datasets
│   ├── MCF-7
│   │   ├── MCF-7_A.txt
│   │	├── MCF-7_graph_indicator.txt
│   │	├── MCF-7_graph_labels.txt
│   │	├── MCF-7_node_labels.txt
│   ├── dataset.py  
│   ├── datautils.py
│   ├── name.py

Use dataset.py to split train, val, and test sets.

Example

python dataset.py --data MCF-7 --trainsz 0.7 --testsz 0.15

Experiments

Parameters

• data: dataset name, default = 'MCF-7'
• lr: learning rate, default = 5e-3
• batchsize: batch size, default = 512
• nepoch: number of training epochs, default = 100
• hdim: hidden dimension of RQGNN, default = 64
• width: width of RQGNN, default = 4
• depth: depth of RQGNN, default = 6
• dropout: dropout rate, default = 0.4
• normalize: batch normalize, default = 1
• beta: hyperparameter in loss function, default = 0.999
• gamma: hyperparameter in loss function, default = 1.5
• decay: weight decay, default = 0
• seed: random seed, default = 10
• patience: patience for training, default = 50

Example

python main.py --data MCF-7 --lr 5e-3 --batchsize 512 --nepoch 100 --hdim 64 --width 4 --depth 6 --dropout 0.4 --normalize 1 --beta 0.999 --gamma 1.5 --decay 0 --seed 10 --patience 50

Citation

@inproceedings{rqgnn,
author = {Xiangyu, Dong and Xingyi, Zhang and Sibo, Wang},
title = {Rayleigh Quotient Graph Neural Networks for Graph-level Anomaly Detection},
year = {2024},
booktitle = {ICLR},
}