Disjoint or Overlapping? Inference Windowing for Reconstruction-Based Time Series Anomaly Detection

In three linesComparative study of reconstruction-based anomaly detection methods for time series. Authors propose unified evaluation protocol on TSB-AD benchmark and show overlapping inference windows improve performance by +28% average across PCA, DLinear, AutoEncoder, TimesNet, and Transformer variants.

Summary generated by Claude — human-verified