Reproduce “Enhancing Graph Neural Network-based Fraud Detectors against Camouflaged Fraudsters” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yingtong Dou, Zhiwei Liu, Li Sun, Yutong Deng, Hao Peng, Philip S. Yu
Researchers from UIC and collaborators developed CARE-GNN, a Graph Neural Network designed to counter two types of fraudster camouflage: feature and relation manipulation. The model achieved superior performance on Yelp and Amazon fraud datasets, outperforming existing GNN-based detectors by adaptively filtering noisy neighbors and learning label-aware node similarities.
Key insights
- Fraudsters exhibit both feature camouflage (manipulating features to appear benign) and relation camouflage (connecting to many benign entities), which directly impacts GNN performance.
- Learning a label-aware similarity measure is crucial for identifying disguised fraudulent features and effectively distinguishing informative from camouflaged neighbors.
- An adaptive neighbor selection mechanism, guided by reinforcement learning, can dynamically determine optimal filtering thresholds, making explicit attention mechanisms less critical for relation-aware aggregation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.