Research map / Time series / Anomaly detection
Graph Mixture of Experts and Memory-augmented Routers for Multivariate Time Series 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 Graph-MoE, a plug-and-play mixture-of-experts over hierarchical GNN layers with memory-augmented routers, for unsupervised multivariate time series anomaly detection.
Problem
Unsupervised multivariate time series anomaly detection with GNN-based methods, which only use the last GNN layer's output and ignore intermediate multi-hop representations.
Approach
Treat each GNN layer as an expert that aggregates entity features with global-guided attention, and use a memory-augmented router (recurrent memory of historical features with gates) to weight the experts' outputs, so intermediate-layer information is integrated adaptively.
Key results
- Graph-MoE reaches 87.2 AUROC on SWaT and 94.2 on WADI, versus 84.8 and 91.9 for MTGFlow.
- Integrating Graph-MoE into GANF raises SWaT AUROC from 79.8 to 82.6; into MTGFlow from 84.8 to 87.2; into USD from 90.2 to 92.3.
- Ablation on SWaT: removing both MoE and MAR gives 85.5; MoE alone 86.8; MAR alone 86.2; both 87.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Graph-MoE instead of MTGFlow | MTS anomaly detection · SWaT, WADI, PSM, MSL, SMD | AUROC: 87.2$\pm$1.3 vs 84.8$\pm$1.5 (SWaT); 94.2$\pm$0.8 vs 91.9$\pm$1.1 (WADI); 88.0$\pm$0.7 vs 85.7$\pm$1.5 (PSM); 72.1$\pm$1 | Helped |
| Vary the number of experts from 1 to 4; 3 experts performs best (ablation) | Graph-MoE · SWaT and WADI | AUROC: SWaT: 86.2$\pm$0.7 (1), 86.8$\pm$0.4 (2), 87.2$\pm$1.3 (3), 85.6$\pm$2.1 (4); WADI: 92.5$\pm$1.2 (1), 93.4$\pm$0.6 (2), | Mixed |
| Add MoE while keeping MAR (ablation) | Graph-MoE · SWaT | AUROC: 86.2$\pm$1.2 -> 87.2$\pm$1.3 | Helped |
| Add MoE while keeping MAR (ablation) | Graph-MoE · WADI | AUROC: 92.9$\pm$1.0 -> 94.2$\pm$0.8 | Helped |
| Add MAR while keeping MoE (ablation) | Graph-MoE · SWaT | AUROC: 86.8$\pm$1.2 -> 87.2$\pm$1.3 | Helped |
| Add MAR while keeping MoE (ablation) | Graph-MoE · WADI | AUROC: 93.5$\pm$0.4 -> 94.2$\pm$0.8 | Helped |
| Increase Graph-MoE layer count up to 3, then increase it to 4 (ablation) | Graph-MoE · Not specified in the excerpt | AUROC | Mixed |
| Add Graph-MoE to GANF | GANF · SWaT | AUROC: 79.8 -> 82.6; gain of 3.5 percentage points | Helped |
| Add Graph-MoE to MTGFlow | MTGFlow · SWaT | AUROC: 84.8 -> 87.2; gain of 2.8% | Helped |
| Add Graph-MoE to USD | USD · SWaT | AUROC: 90.2 -> 92.3; gain of 2.3 percentage points | Helped |
Limitations
- Evaluation limited to window-level AUROC on five datasets
- Number of experts and layers tuned by ablation on SWaT and WADI only
Experimental setup
Datasets: SWaT WADI PSM MSL SMD
Benchmarks: SWaT WADI PSM MSL SMD
Metrics: AUROC
Baselines: DeepSVDD DAGMM ALOCC DROCC DeepSAD USAD GANF MTGFlow USD
Builds on: MTGFlow GANF mixture of experts normalizing flow
Closest papers by meaning
- Multi-Scale Dynamic Graph Learning for Time Series Anomaly Detection (Student Abstract) (AAAI 2024 · similarity 0.89)
- The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.87)
- Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders (ICLR 2025 · similarity 0.86)
- A Generalizable Anomaly Detection Method in Dynamic Graphs (AAAI 2025 · similarity 0.86)
- UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.86)
- GCAD: Anomaly Detection in Multivariate Time Series from the Perspective of Granger Causality (AAAI 2025 · 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.