Research map / Graph learning / Graph self-supervised learning

Revisiting Graph-Based Fraud Detection in Sight of Heterophily and Spectrum

AAAI 2024 · Graph learning › Graph self-supervised learning · approach: Graph anomaly detection · 57 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SEC-GFD, a semi-supervised GNN fraud detector that combines a spectrum-based hybrid-pass filter (band-pass wavelet bands plus high-pass bands of low-order neighbors) with a local environmental constraint contrastive loss, evaluated on four fraud datasets.

Problem

Graph-based fraud detection is a semi-supervised node classification task on heterophilic, class-imbalanced graphs, where vanilla GNNs (low-pass filters) perform poorly and existing models under-use node label information.

Approach

Decompose the graph spectrum into mixed-frequency bands (BWGNN-style beta wavelet band-pass filters plus high-pass filters of low-order neighbors, aggregated together), and add a contrastive local-environment constraint that pulls normal nodes' multi-hop neighbor representations toward their kNN feature-view representations while pushing anomalous nodes apart.

Key results

• SEC-GFD reports F1-macro/AUC of 77.73/91.39 on YelpChi, 92.35/98.23 on Amazon, 89.86/96.32 on T-Finance and 87.74/96.11 on T-Social, best on all four datasets versus listed baselines.
• Ablation on Amazon and T-Finance shows removing band-pass bands causes the largest drop (F1-macro 92.27 to 90.29 on Amazon), while removing the local environmental constraint drops F1-macro by 1.31 points on Amazon.
• Heterophily-edge clipping experiments show that removing heterophily edges in the training graph does not reliably help, and band-pass filters perform best overall.

Measured effects reported in the paper

Surprising findings

• Clipping heterophily edges on the whole graph helps low-pass filters, but clipping only training-set heterophily edges leaves low-pass filters poor and degrades band-pass and high-pass filters, suggesting indiscriminate heterophily edge removal is not necessarily beneficial.
• Performance on T-Social keeps improving until neighbor order C reaches 5, whereas on the three smaller datasets gains stop after C=2.

Limitations

• Hyperparameters such as C and the hidden dimension differ per dataset and are tuned, with C=5 needed for T-Social.
• Experiments are limited to four datasets and static node-level fraud detection.
• The local environmental constraint uses kNN graphs built from features, adding a construction step whose cost and sensitivity are not analyzed in the text.

Experimental setup

Datasets: Amazon YelpChi T-Finance T-Social

Benchmarks: YelpChi Amazon T-Finance T-Social

Metrics: F1-macro AUC

Baselines: GCN GraphSAGE GAT GIN GraphConsis CARE-GNN PC-GNN GAGA ACM H2-FDetector BWGNN GDN

Builds on: BWGNN InfoNCE SGC GCN

Closest papers by meaning

• Partitioning Message Passing for Graph Fraud Detection (ICLR 2024 · similarity 0.93)
• A Label-free Heterophily-guided Approach for Unsupervised Graph Fraud Detection (AAAI 2025 · similarity 0.91)
• Dynamic Neighborhood Modeling via Node-Subgraph Contrastive Learning for Graph-Based Fraud Detection (AAAI 2025 · similarity 0.91)
• DiG-In-GNN: Discriminative Feature Guided GNN-Based Fraud Detector against Inconsistencies in Multi-Relation Fraud Graph (AAAI 2024 · similarity 0.90)
• Barely Supervised Learning for Graph-Based Fraud Detection (AAAI 2024 · similarity 0.89)
• Context-aware Graph Neural Network for Graph-based Fraud Detection with Extremely Limited Labels (AAAI 2025 · similarity 0.86)