Research map / Graph learning / Graph self-supervised learning
Dynamic Spectral Graph Anomaly Detection
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 DSGAD, a spectral GNN for graph anomaly detection that learns trainable Beta-mixture wavelets and fuses wavelet-specific features via energy-difference-aware MLP extractors and channel convolution (Conv1D) fusion.
Problem
Spectral GNNs for graph anomaly detection rely on hand-designed graph wavelets and sequential feature concatenation, which capture incomplete anomaly information and fuse features inconsistently.
Approach
Replace fixed Beta wavelets with trainable Beta-mixture wavelets whose mixing weights depend on node attributes (dynamic wavelets), and fuse wavelet-specific features by per-wavelet MLP extractors followed by channel-concatenation and Conv1D fusion that preserves location correlations.
Key results
- DSGAD (homo) averages 89.63 AUC and 80.67 F1-macro across four homogeneous-graph datasets, versus 88.16 / 78.07 for the best baseline GHRN (homo).
- DSGAD (hete) reaches 93.67 average AUC and 84.11 F1-macro on YelpChi and Amazon, above GHRN (hete) at 93.04 / 83.26.
- Ablation shows removing dynamic wavelets, wavelet feature extraction and channel convolution fusion together drops average AUC by 1.78 points and F1-macro by 3.43 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DSGAD instead of GPR-GNN | Homogeneous graph anomaly detection · YelpChi | AUC: 82.98±0.62 -> 84.77±0.28 | Helped |
| use DSGAD instead of BWGNN | Homogeneous graph anomaly detection · YelpChi | F1-macro: 68.41±0.57 -> 71.86±0.24 | Helped |
| use DSGAD instead of GHRN | Homogeneous graph anomaly detection · T-finance | AUC: 94.92±0.50 -> 96.51±0.15 | Helped |
| use DSGAD instead of GPR-GNN | Homogeneous graph anomaly detection · T-finance | F1-macro: 88.04±4.12 -> 91.80±0.41 | Helped |
| use DSGAD instead of GPR-GNN | Homogeneous graph anomaly detection · Amazon | AUC: 97.79±1.08 -> 97.91±0.17 | Helped |
| use DSGAD instead of GPR-GNN | Homogeneous graph anomaly detection · Amazon | F1-macro: 93.06±0.38 -> 92.32±0.22 | Hurt |
| use DSGAD instead of GHRN | Homogeneous graph anomaly detection · Tolokers | AUC: 79.10±0.28 -> 79.31±0.48 | Helped |
| use DSGAD instead of GHRN | Homogeneous graph anomaly detection · Tolokers | F1-macro: 66.60±0.47 -> 66.70±0.67 | Helped |
| use DSGAD instead of GHRN | Heterogeneous graph anomaly detection · YelpChi | AUC: 88.97±0.48 -> 89.30±0.58 | Helped |
| use DSGAD instead of GHRN | Heterogeneous graph anomaly detection · YelpChi | F1-macro: 75.09±0.59 -> 75.49±0.75 | Helped |
| use DSGAD instead of GHRN | Heterogeneous graph anomaly detection · Amazon | AUC: 97.11±0.46 -> 97.73±0.26 | Helped |
| use DSGAD instead of BWGNN | Heterogeneous graph anomaly detection · Amazon | F1-macro: 91.82±0.55 -> 92.14±0.44 | Helped |
Surprising findings
- Ablation shows that removing the channel convolution fusion alone still leaves average AUC within about 1 point of the full model, so most of the gain comes from the other components.
Limitations
- Dynamic wavelet computation may be expensive on large-scale graphs due to high resource demands.
- Only structural (frequency-domain) anomalies are modeled; attribute anomalies are not handled.
Experimental setup
Datasets: T-finance Tolokers YelpChi Amazon
Benchmarks: T-finance Tolokers YelpChi Amazon
Metrics: AUC F1-macro
Baselines: GCN GAT GIN ChebyNet CARE-GNN GPR-GNN BWGNN GHRN
Builds on: BWGNN GHRN Beta wavelets
Closest papers by meaning
- A Generalizable Anomaly Detection Method in Dynamic Graphs (AAAI 2025 · similarity 0.88)
- Rayleigh Quotient Graph Neural Networks for Graph-level Anomaly Detection (ICLR 2024 · similarity 0.88)
- UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.87)
- Graph Anomaly Detection as Dynamical Transport: Training-Free Scoring via Empirical Bayes (NeurIPS 2026 · similarity 0.86)
- Self-Discriminative Modeling for Anomalous Graph Detection (ICML 2025 · similarity 0.86)
- ARC: A Generalist Graph Anomaly Detector with In-Context Learning (NeurIPS 2024 · similarity 0.86)
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.