Editor: Multi-Resolution Cleaning of Multivariate Time Series Via Detect-Localize-Repair

Multivariate time series (MTS) power downstream decisions, yet real-world MTS often contain errors with varying magnitudes and granularities (e.g., points, subsequences, crossvariable), breaking temporal and inter-variable dependencies. Most existing cleaning methods rely on fixed windows, which (i) miss subtle deviations and (ii) over-clean correct values. We introduce EDITOR, a multi-resolution framework that separates error detection, localization, and repair to improve cleaning precision while preserving MTS structure: (i) High-Sensitivity Detection (HSD) identifies erroneous windows (coarse regions likely to contain errors) using a dual-attention UNet, a differenceenhanced UNet, and dynamic thresholding to capture a broad range of anomalies. (ii) Multi-Granularity Localization (MGL) operates within each detected window, casting fine-grained localization as constrained optimization under learned temporal and cross-variable dependencies. It leverages evolutionary search to pinpoint erroneous points, subsequences or cross-variable concurrence while minimizing collateral changes. (iii) ContextAware Repair (CAR) applies a two-stage bidirectional correction: a Temporal Convolutional Network (TCN) restores temporal coherence, followed by a Graph Convolutional Network (GCN) for cross-variable consistency, both in forward and backward directions. EDITOR outperforms strong baselines across five datasets and improves downstream tasks. Ablation studies confirm the necessity of each module.