G lassbox AD: An Interactive System for Dissecting Hierarchical Time-Series Anomaly Detection
Mingyi Huang, Qinghua Liu, Paul Boniol, John Paparrizos · 2026
Time-series anomaly detection (TSAD) is challenging in unsupervised settings because anomalies are heterogeneous and often manifest at different temporal scales. HYDRA addresses this by constructing a multi-level hierarchy of representative subsequences, computing reference-driven anomaly scores at each level, and aggregating them into a final score. In this paper, we present GlassboxAD, an interactive demo system that makes HYDRA's hierarchical behavior observable and interpretable. The system supports two entry points: instant browsing of benchmark results on TSB-AD and on-demand analysis of user-uploaded time series. Users can inspect how detections emerge across levels and assess robustness under different tolerance and threshold settings through linked views: layered visualizations of representatives, per-layer evidence for selected subsequences, score-layer switching on the timeline, and specific evaluation metrics for TSAD. We further contextualize these interactive diagnoses with benchmark-driven comparisons, showing that HYDRA attains a top overall rank on TSB-AD while enabling users to connect case-level, hierarchy-based explanations with suite-level trends across heterogeneous anomaly regimes. The webpage is available online: https://glassboxad.streamlit.app/