Research map / Time series / Anomaly detection

ScatterAD: Temporal-Topological Scattering Mechanism for Time Series Anomaly Detection

NeurIPS 2025 · Time series › Anomaly detection · approach: Time series anomaly detection evaluation · 1 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 ScatterAD, a dual-encoder (temporal and topological on a temporal graph) time series anomaly detector that pulls normal representations toward a scatter center on the hypersphere, with a time-consistency loss, EMA target encoder and contrastive fusion, evaluated on six IoT benchmarks.

Problem

Multivariate time series anomaly detection for industrial IoT, where existing methods model temporal or spatial (topological) dependencies separately and are insensitive to anomalous dispersion of representations in high-dimensional space.

Approach

Treats each time step as a graph node, models representation "scattering" on the unit hypersphere with a scatter-center loss plus a temporal smoothness constraint, and fuses online/target encoder views via contrastive alignment, motivated by an information-bottleneck view of temporal-topological complementarity.

Key results

• On MSL, ScatterAD reaches Aff-F 0.867 and PA-F 0.964, the best among compared methods in the main table.
• ScatterAD ranks first in 15 of 24 main-table settings and achieves the highest Affiliated-F1 on five of six datasets, but its Aff-F on NIPS-TS-SWAN is only 0.038 versus 0.805 for Sub-Adjacent Transformer.
• Ablations show removing the contrastive fusion module degrades AF on SWaT from 0.704 to 0.701 and on WADI from 0.587 to 0.549.

Measured effects reported in the paper

Surprising findings

• On NIPS-TS-SWAN, ScatterAD has a very low Affiliated-F1 (0.038) and Affiliated recall (0.019) despite top PA-F and AUC scores.
• The separation ratio of anomalous to normal scattering in the noise analysis falls below 1.0 for WADI at sigma=1.0 and 2.0 (0.83 and 0.82), contradicting the claimed universality of the scattering phenomenon.

Limitations

• Dependence on graph construction quality and assumption of static topology, as stated in the checklist.
• Lack of explicit causality modeling, as stated in the checklist.
• Mixed results on NIPS-TS-SWAN, with very low affiliated F1.

Experimental setup

Datasets: MSL PSM SWaT WADI NIPS-TS-SWAN NIPS-TS-GECCO

Benchmarks: MSL PSM SWaT WADI NIPS-TS-SWAN NIPS-TS-GECCO

Metrics: Affiliated-F1 PA-F1 AUC-ROC AUC-PR Range-AUC-ROC Range-AUC-PR VUS-ROC VUS-PR Affiliated precision Affiliated recall

Baselines: Sub-Adjacent Transformer TopoGDN MEMTO DuoGAT MTGFlow iTransformer DCdetector Anomaly Transformer GANF ModernTCN VAE Isolation Forest

Builds on: DCdetector GAT information bottleneck Deep InfoMax

Closest papers by meaning

• Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.88)
• Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability. (NeurIPS 2026 · similarity 0.87)
• DiffGAD: A Diffusion-based Unsupervised Graph Anomaly Detector (ICLR 2025 · similarity 0.87)
• Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.86)
• The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.86)
• SARAD: Spatial Association-Aware Anomaly Detection and Diagnosis for Multivariate Time Series (NeurIPS 2024 · similarity 0.86)