UniGAD: Unifying Multi-level Graph Anomaly Detection

This is the official implementation of the following paper:

UniGAD: Unifying Multi-level Graph Anomaly Detection
Yiqing Lin, Jianheng Tang, Chenyi Zi, H.Vicky Zhao, Yuan Yao, Jia Li
NeurIPS 2024

Environment Setup

Before you start, install Pytorch and DGL with appropriate CUDA support. Please refer to the PyTorch and DGL websites for the specific guidelines.

Install dependencies:

pip install -r requirements.txt

Datasets

We release a real-world large-scale social group dataset T-Group in the paper, which contains both node anomaly labels and graph anomaly labels. The dataset is on google drive. Download and unzip the file in the dataset folder.

Experiments

Parameters:

• datasets: dataset ID
• batch_size: batch size of graphs
• cross_mode: cross traning mode for mult-level tasks, default = "ng2ng"
• khop: MRQSampler tree depth, default = 1
• kernels: the model used in the pretrained encoder, default = gcn
• epoch_pretrain: number of pretrain epochs, default = 50
• lr: learning rate of the pretraning, default = 5e-4
• epoch_ft: number of GraphStitch Network traning epochs, default = 200
• lr_ft: learning rate of the GraphStitch Network, default = 0.003
• trials: number of experiment repetitions, default = 1

Example:

mkdir results/
cd src/
# reproduce the results on dataset weibo
python main.py --datasets 1 --pretrain_model graphmae --kernels bwgnn,gcn --lr 5e-4 --save_model --epoch_pretrain 50 --batch_size 1 --khop 1 --epoch_ft 300 --lr_ft 0.003 --final_mlp_layers 3 --cross_modes ne2ne,n2ne,e2ne --metric AUROC --trials 5

Note: In our experiments, we found that typically hop = 1 or 2 is optimal, so the MRQSampler in our code is implemented as two unfolded versions dedicated to orders 1 and 2 rather than a recursive version for the performance.

Reference

Encoder Model

GraphMAE (https://github.com/THUDM/GraphMAE).

Pareto loss function

Multi-task loss optimizer (https://github.com/tomtang110/multi-task_loss_optimizer/tree/master)

Baseline Information

Node and Edge level Baselines (https://github.com/squareRoot3/GADBench/).

OCGIN,OCGTL (https://github.com/boschresearch/GraphLevel-AnomalyDetection).

GLocalKD (https://github.com/RongrongMa/GLocalKD).

iGAD (https://github.com/graph-level-anomalies/iGAD/tree/main).

GmapAD (https://github.com/XiaoxiaoMa-MQ/GmapAD).

RQGNN (https://github.com/xydong127/RQGNN/tree/main).

GraphPrompt (https://github.com/Starlien95/GraphPrompt)

All in One (https://github.com/sheldonresearch/ProG)