Research map / Graph learning / Graph self-supervised learning
Semi-supervised Graph Anomaly Detection via Robust Homophily Learning
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 RHO, a semi-supervised graph anomaly detector that learns adaptive spectral filters (AdaFreq) in cross-channel and channel-wise views and aligns them with a contrastive loss (GNA) to capture diverse homophily among labeled normal nodes.
Problem
Semi-supervised graph anomaly detection assumes labeled normal nodes share similar homophily and represent the whole normal class, but normal nodes in real GAD datasets exhibit diverse homophily, causing low-homophily normal nodes to be misclassified as anomalies.
Approach
Replace fixed graph filters with learnable frequency-response filters (g(λ)=1−kλ) applied both across channels (one shared k) and per feature channel (d learnable k's), train them with a one-class hypersphere loss on labeled normal nodes, and align the two views' representations with a contrastive GNA loss.
Key results
- RHO achieves AUROC 0.8509 on Elliptic and 0.8623 on T-Finance, best among compared methods on six datasets, with up to 12.19% AUROC gain over GGAD.
- GGAD gets the highest AUROC on Amazon (0.9443 vs RHO 0.9302) and Reddit (0.6354 vs RHO 0.6207).
- Removing the GNA alignment loss consistently lowers both AUROC and AUPRC in the ablation study.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GNA alignment loss to the cross-channel and channel-wise one-class losses (ablation) | RHO · Eight GAD datasets | AUROC: 0.6117 -> 0.6207 (Reddit); 0.5727 -> 0.6255 (Tolokers); 0.6361 -> 0.7129 (Photo); 0.8692 -> 0.9302 (Amazon); 0.7954 -> 0 | Helped |
| Add GNA alignment loss to the cross-channel and channel-wise one-class losses (ablation) | RHO · Eight GAD datasets | AUPRC: 0.0556 -> 0.0616 (Reddit); 0.3024 -> 0.3256 (Tolokers); 0.1398 -> 0.2337 (Photo); 0.5443 -> 0.7879 (Amazon); 0.2657 -> 0 | Helped |
| Add the channel-wise view to the cross-channel view (ablation) | RHO · Eight GAD datasets | AUROC: 0.5756 -> 0.6117 (Reddit); 0.5712 -> 0.5727 (Tolokers); 0.6322 -> 0.6361 (Photo); 0.7436 -> 0.8692 (Amazon); 0.7371 -> 0 | Mixed |
| Add the cross-channel view to the channel-wise view (ablation) | RHO · Eight GAD datasets | AUROC: 0.6095 -> 0.6117 (Reddit); 0.5878 -> 0.5727 (Tolokers); 0.6144 -> 0.6361 (Photo); 0.8905 -> 0.8692 (Amazon); 0.7718 -> 0 | Mixed |
| Add the channel-wise view to the cross-channel view (ablation) | RHO · Eight GAD datasets | AUPRC: 0.0431 -> 0.0556 (Reddit); 0.3052 -> 0.3024 (Tolokers); 0.1390 -> 0.1398 (Photo); 0.2423 -> 0.5443 (Amazon); 0.1862 -> 0 | Mixed |
| Add the cross-channel view to the channel-wise view (ablation) | RHO · Eight GAD datasets | AUPRC: 0.0481 -> 0.0556 (Reddit); 0.2923 -> 0.3024 (Tolokers); 0.1451 -> 0.1398 (Photo); 0.5895 -> 0.5443 (Amazon); 0.2287 -> 0 | Mixed |
| Increase the GNA weight alpha | RHO · Elliptic, Tolokers, Amazon, Question, and T-Finance | AUROC and AUPRC | Helped |
| Increase the GNA weight alpha | RHO · Reddit and Photo | AUROC and AUPRC | Hurt |
| Increase the number of labeled normal training nodes | RHO · GAD datasets | AUROC | Helped |
Surprising findings
- GGAD, a generative method, outperforms RHO on Reddit and Amazon in AUROC, which the authors attribute to generated anomalies not generalizing across datasets.
- Fully supervised methods PMP and CONSISGAD perform poorly when adapted to the semi-supervised setting.
Limitations
- Transductive method; cannot directly handle inductive settings.
- Sensitive to learning rate and alpha, which need dataset-specific tuning (e.g., lr 5e-6 for DGraph, alpha 0.1 for Reddit and Photo).
- Theorem relies on orthogonality assumptions and dropped cross terms, and assumes center c=1, so the closed-form is approximate.
Experimental setup
Datasets: Reddit Tolokers Photo Amazon Elliptic Question T-Finance DGraph
Benchmarks: Reddit Tolokers Photo Amazon Elliptic Question T-Finance DGraph
Metrics: AUROC AUPRC
Baselines: DOMINANT AnomalyDAE OCGNN AEGIS GAAN TAM GGAD GCN BWGNN PMP CONSISGAD
Builds on: GCN BWGNN GGAD one-class hypersphere objective (OCGNN)
Closest papers by meaning
- Generative Semi-supervised Graph Anomaly Detection (NeurIPS 2024 · similarity 0.89)
- ARC: A Generalist Graph Anomaly Detector with In-Context Learning (NeurIPS 2024 · similarity 0.88)
- Open-Set Graph Anomaly Detection via Normal Structure Regularisation (ICLR 2025 · similarity 0.88)
- Leveraging Diffusion Model as Pseudo-Anomalous Graph Generator for Graph-Level Anomaly Detection (ICML 2025 · similarity 0.87)
- Homophily Enhanced Graph Domain Adaptation (ICML 2025 · similarity 0.87)
- ADA-GAD: Anomaly-Denoised Autoencoders for Graph Anomaly Detection (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.