Research map / Time series / Anomaly detection

CrossAD: Time Series Anomaly Detection with Cross-scale Associations and Cross-window Modeling

NeurIPS 2025 · Time series › Anomaly detection · approach: Time series anomaly detection evaluation · 3 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 CrossAD, which reconstructs finer-grained series from coarser ones to model cross-scale associations, plus a learnable query library and global multi-scale context to go beyond sliding-window limits, for unsupervised time series anomaly detection.

Problem

Unsupervised time series anomaly detection with multi-scale modeling: existing methods model scales independently or fuse features simply, ignoring cross-scale associations, and their fixed sliding windows limit access to global context.

Approach

Cross-scale reconstruction trains a Transformer to reconstruct each finer-grained scale from coarser scales via a cross-scale mask, using reconstruction failure under disrupted cross-scale association as the anomaly signal; a period-aware Gumbel-Softmax router selects learnable sub-series queries, and an EMA-updated global prototype bank injects cross-window context into decoding.

Key results

• On seven datasets, CrossAD reports the best VUS-PR on all of them, e.g. 0.6211 on GECCO versus 0.4819 for ModernTCN and 0.5596 on PSM versus 0.4689 for ModernTCN.
• Ablation shows multi-scale reconstruction raises average VUS-PR by 8.13% over the single-scale baseline, and cross-scale reconstruction adds a further 9.18% VUS-PR gain on top.
• CrossAD uses 0.93M parameters versus 4.68M for TimesNet, with inference time of 0.80 s versus 1.68 s on GECCO.

Measured effects reported in the paper

Surprising findings

• Anomaly Transformer scores VUS-PR 0.0278 on GECCO, far below its VUS-ROC, suggesting its scores are poorly calibrated under this protocol.

Limitations

• Uses one unified hyperparameter set across datasets, while optimal numbers of queries and prototypes vary by dataset.
• Cross-scale and cross-window gains are shown mainly through ablations on a few datasets rather than across all datasets.

Experimental setup

Datasets: SMD MSL SMAP SWaT PSM GECCO SWAN

Benchmarks: SMD MSL SMAP SWaT PSM GECCO SWAN

Metrics: VUS-PR VUS-ROC AUC-ROC AUC-PR Standard-F1 Affiliation-F1 Range-AUC-ROC Range-AUC-PR Accuracy

Baselines: OCSVM PCA IForest LODA HBOS LOF AutoEncoder DAGMM LSTM CAE-Ensemble OmniAnomaly Anomaly Transformer

Builds on: TimesNet Anomaly Transformer DCdetector SPOT

Closest papers by meaning

• Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.90)
• Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability. (NeurIPS 2026 · similarity 0.88)
• Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection (NeurIPS 2026 · similarity 0.87)
• The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.87)
• Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.87)
• SARAD: Spatial Association-Aware Anomaly Detection and Diagnosis for Multivariate Time Series (NeurIPS 2024 · similarity 0.86)