HUGE: A Label-Free Heterophily-Guided Approach for Unsupervised Graph Fraud Detection
May 19, 2026 ยท View on GitHub
Junjun Pan, Yixin Liu, Xin Zheng, Yizhen Zheng, Alan Wee-Chung Liew, Fuyi Li, Shirui Pan
This repo contains the official implementation of AAAI25 HUGE: A Label-Free Heterophily-Guided Approach for Unsupervised Graph Fraud Detection
To reproduce the results proposed in the paper, run
Amazon
python main.py --dataset Amazon --lr 5e-4 --kd_param 0.5 --epoch 300 --seed 0
python main.py --dataset Facebook --lr 5e-4 --kd_param 0.5 --epoch 300 --seed 0
python main.py --dataset Reddit --lr 5e-4 --kd_param 0.5 --epoch 300 --seed 0
YelpChi
python main.py --dataset YelpChi --lr 5e-4 --kd_param 0.5 --epoch 300 --seed 0
AmazonFull
python main.py --dataset AmazonFull --lr 5e-4 --kd_param 1.0 --epoch 10 --seed 0
YelpChiFull
python main.py --dataset YelpChiFull --lr 1e-5 --kd_param 3 --epoch 5 --seed 0
Environment
The code is tested under conda environment (py 3.9.18) with these additional libs installed:
Please let us know if you find other libs are also required.
dgl==1.1.2+cu117
torch==2.0.0+cu117
torch-geometric==2.5.2+cu117
torch-cluster==1.6.3+cu117
torch-scatter==2.1.2+cu117
torch-sparse==0.6.18+cu117
tqdm==4.64.1