Research map / Time series / Anomaly detection

Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders

ICLR 2025 · Time series › Anomaly detection · approach: Time series anomaly detection evaluation · 0 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.

Pre-trains a general time series anomaly detector (DADA) on multi-domain data with adaptive bottleneck selection and dual adversarial decoders, evaluated zero-shot on downstream anomaly detection benchmarks.

Problem

Time series anomaly detection models are trained per dataset and generalize poorly across domains; a single general model must handle varying information-bottleneck needs and distinguish normal from diverse anomalies without target-domain training.

Approach

Combines a pool of different-sized bottlenecks chosen per series by a noisy top-k adaptive router with dual decoders: a normal decoder for normal reconstruction and an anomaly decoder trained adversarially (via gradient reversal) on injected anomalies, with complementary-mask reconstruction and variance-based anomaly scores.

Key results

• Zero-shot DADA achieves affiliated F1 of 84.57 on SMD, 78.48 on MSL, 75.42 on SMAP, 74.60 on SWaT, 82.26 on PSM, best among listed baselines trained per dataset.
• On NeurIPS-TS, DADA reports best F1 on CICIDS (73.49), Creditcard (75.12), and SWAN (71.93) and second on GECCO (90.20 F1 vs D3R 91.83).
• Ablation: removing AdaBN drops average F1 from 79.49 to 74.45; removing adversarial training and dual decoders drops to 63.23 average F1.

Measured effects reported in the paper

Surprising findings

• Training without any human labels (anomaly injection only) yields nearly identical F1 (e.g., 84.69 vs 84.57 on SMD).
• Replacing adaptive bottlenecks with a mixture-of-experts of same-size bottlenecks (MoE 256) underperforms AdaBN on several datasets such as SWaT, GECCO, and Creditcard.

Limitations

• Evaluation relies on affiliated F1 and AUC; PA-based comparisons with baselines are not directly reported in the main text.
• Anomaly injection defines the anomaly types used in adversarial training, so unseen anomaly types may be missed unless added to pre-training, as the authors note.

Experimental setup

Datasets: Monash ASD Exathlon ECG MITDB OPP SVDB GAIA IOPS MGAB NYC SKAB

Benchmarks: SMD MSL SMAP SWaT PSM NeurIPS-TS (CICIDS, Creditcard, GECCO, SWAN) UCR

Metrics: affiliated F1 affiliated precision affiliated recall AUC_ROC VUS F1

Baselines: OCSVM PCA IForest LODA AutoEncoder DAGMM BeatGAN OmniAnomaly Anomaly Transformer MEMTO DCdetector D3R

Builds on: GPT4TS Anomaly Transformer gradient reversal layer SPOT

Closest papers by meaning

• The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.89)
• Fortifying Time Series: DTW-Certified Robust Anomaly Detection (NeurIPS 2025 · similarity 0.87)
• Position: Quo Vadis, Unsupervised Time Series Anomaly Detection? (ICML 2024 · similarity 0.87)
• Graph Mixture of Experts and Memory-augmented Routers for Multivariate Time Series Anomaly Detection (AAAI 2025 · similarity 0.86)
• Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection (NeurIPS 2026 · similarity 0.86)
• Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.86)