Research map / Time series / Anomaly detection

VACE: Learning Geometrically Structured Representations for Time Series Anomaly Detection

NeurIPS 2026 · Time series › Anomaly detection · approach: Multivariate anomaly detection · 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.

Proposes VACE, a self-supervised multivariate time-series anomaly detector that trains a channel-aware 1D-CNN encoder with a velocity-consistency loss and scores anomalies by multiplying a Mahalanobis positional score with a velocity-bank directional score.

Problem

Multivariate time series anomaly detection with no labels, where distance-based scoring needs an embedding space with compact, directionally coherent normal structure, which contrastive pair-sampling does not explicitly provide.

Approach

Replace contrastive pair sampling with a negative-free velocity-consistency pretext loss that makes normal embedding trajectories locally piecewise linear, combined with a channel-aware depthwise-separable encoder and a multiplicative positional-times-directional anomaly score.

Key results

• On the 180-series TSB-AD-M evaluation set, VACE reaches VUS-PR 0.48 vs 0.43 for PaAno (previous self-supervised SOTA) and 0.41 for KAN-AD.
• VACE obtains the highest AUC-PR (0.43), F1 (0.46) and Range-F1 (0.46) in the main table, while VUS-ROC (0.78) and AUC-ROC (0.76) are comparable to PaAno.
• Ablation: removing the channel-aware encoder drops VUS-PR by 0.039 and removing the velocity pretext drops it by 0.037; removing velocity scoring costs only 0.006.

Surprising findings

• Velocity scoring contributes only a small drop (0.006 VUS-PR) relative to the pretext, suggesting the directional signal is partly redundant with the positional one.
• Removing BatchNormalisation causes embedding collapse (effective rank 0.025), showing the scorer is sensitive to the geometry regime.

Limitations

• Single-Gaussian Mahalanobis fit may be a noisier approximation of the normal region for very long training series.
• Channel-independent-then-mix encoder may discard inter-channel relational structure needed for relational anomalies.
• Reconstruction-based xLSTM-AD outperforms VACE on several categories (Genesis, SWaT, OPPORTUNITY, PSM).

Experimental setup

Datasets: TSB-AD-M

Benchmarks: TSB-AD-M

Metrics: VUS-PR VUS-ROC AUC-ROC AUC-PR F1 Range-F1

Baselines: PaAno xLSTMAD KAN-AD DeepAnT CrossAD CATCH TimesNet AnomalyTransformer DCDetector PCA IForest OmniAnomaly

Builds on: PaAno RECURVE ModernTCN RevIN

Closest papers by meaning

• Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.88)
• The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.87)
• Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection (NeurIPS 2026 · similarity 0.86)
• Position: Quo Vadis, Unsupervised Time Series Anomaly Detection? (ICML 2024 · similarity 0.85)
• COMET: Codebook-based Online-adaptive Multi-scale Embedding for Time-series Anomaly Detection (NeurIPS 2026 · similarity 0.85)
• Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.85)