Research map / Graph learning / Graph self-supervised learning
Federated Graph Anomaly Detection Through Contrastive Learning with Global Negative Pairs
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 FedCLGN, a federated graph anomaly detection framework that maintains a server-side pool of global negative pairs (embeddings of pseudo-labeled anomalous nodes uploaded by clients), diffuses them with personalized PageRank, and uses CoLA-style contrastive learning with FedPub-style personalized aggregation.
Problem
Anomaly detection on attributed graphs distributed across clients in federated settings, where restricted data access and missing cross-client edges make contrastive negative-pair sampling less effective and centralized methods degrade.
Approach
Clients pseudo-label anomalous nodes and upload only the negative-pair embeddings of those nodes; the server aggregates them into a global negative pair pool, refines it with personalized PageRank diffusion, and redistributes it to clients to strengthen contrastive positive/negative discrimination, combined with personalized aggregation weighted by negative-pair similarity.
Key results
- FedCLGN achieves the best AUC on all four datasets, e.g., 0.861 on Cora with 5 clients vs 0.806 for the best federated baseline (FedPub-CoLA).
- On Flickr with 5 clients FedCLGN reaches 0.876 AUC vs 0.759 for CoLA and 0.756 for FedAvg-CoLA.
- Parameter study suggests confidence threshold lambda in [-0.3, -0.5] works best, and transfer probability alpha gives stable performance across 0.1-0.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedCLGN instead of the strongest reported baseline | Federated graph anomaly detection · Cora, 5/10/15 clients | AUC: 0.806±0.011 -> 0.861±0.009; 0.796±0.015 -> 0.807±0.010; 0.789±0.011 -> 0.800±0.002 | Helped |
| Use FedCLGN instead of the strongest reported baseline | Federated graph anomaly detection · Citeseer, 5/10/15 clients | AUC: 0.814±0.014 -> 0.840±0.008; 0.814±0.013 -> 0.826±0.008; 0.738±0.003 -> 0.812±0.007 | Helped |
| Use FedCLGN instead of the strongest reported baseline | Federated graph anomaly detection · BlogCatalog, 5/10/15 clients | AUC: 0.825±0.012 -> 0.846±0.004; 0.800±0.008 -> 0.823±0.005; 0.805±0.007 -> 0.840±0.004 | Helped |
| Use FedCLGN instead of the strongest reported baseline | Federated graph anomaly detection · Flickr, 5/10/15 clients | AUC: 0.775±0.006 -> 0.876±0.001; 0.758±0.005 -> 0.858±0.002; 0.806±0.002 -> 0.849±0.009 | Helped |
| Add pseudo-label discovery and global negative-pair diffusion (ablation) | FedPub-CoLA / FedCLGN · Cora, 5/10/15 clients | AUC: 0.806±0.011 -> 0.861±0.009; 0.751±0.009 -> 0.807±0.010; 0.655±0.021 -> 0.800±0.002 | Helped |
| Add pseudo-label discovery and global negative-pair diffusion (ablation) | FedPub-CoLA / FedCLGN · Citeseer, 5/10/15 clients | AUC: 0.796±0.011 -> 0.840±0.008; 0.795±0.003 -> 0.826±0.008; 0.712±0.005 -> 0.812±0.007 | Helped |
| Add pseudo-label discovery and global negative-pair diffusion (ablation) | FedPub-CoLA / FedCLGN · BlogCatalog, 5/10/15 clients | AUC: 0.823±0.004 -> 0.846±0.004; 0.793±0.008 -> 0.823±0.005; 0.805±0.007 -> 0.840±0.004 | Helped |
| Add pseudo-label discovery and global negative-pair diffusion (ablation) | FedPub-CoLA / FedCLGN · Flickr, 5/10/15 clients | AUC: 0.775±0.006 -> 0.876±0.001; 0.722±0.015 -> 0.858±0.002; 0.683±0.017 -> 0.849±0.009 | Helped |
Surprising findings
- Several federated CoLA variants degrade substantially as the number of clients grows from 5 to 15 (e.g., CoLA on Cora drops from 0.728 to 0.625), and FedAvg/FedPub variants fall below DOMINANT/MGAD at 15 clients on Cora.
Limitations
- Anomalies are synthetically injected rather than real ground-truth anomalies
- Performance of several federated variants degrades with more clients and FedCLGN also drops at 15 clients on some datasets
- Uses pseudo-labels with a confidence threshold, so errors in pseudo-labeling may propagate into the negative pool
Experimental setup
Datasets: Cora Citeseer BlogCatalog Flickr
Benchmarks: Cora Citeseer BlogCatalog Flickr
Metrics: ROC-AUC
Baselines: DOMINANT AnomalyDAE MGAD CoLA FedAvg-CoLA FedPub-CoLA
Builds on: CoLA FedPub FedAvg personalized PageRank
Closest papers by meaning
- UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.88)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.87)
- Decoupled Subgraph Federated Learning (ICLR 2025 · similarity 0.87)
- Revisiting Graph Contrastive Learning on Anomaly Detection: A Structural Imbalance Perspective (AAAI 2025 · similarity 0.86)
- ARC: A Generalist Graph Anomaly Detector with In-Context Learning (NeurIPS 2024 · similarity 0.85)
- Graph Anomaly Detection as Dynamical Transport: Training-Free Scoring via Empirical Bayes (NeurIPS 2026 · similarity 0.85)
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.