ProtoX-AD makes time series anomaly detection explainable with prototypes

A team from UiT The Arctic University of Norway (Aitor Sánchez-Ferrera, Elisabeth Wetzer, Kristoffer Wickstrøm, Michael Kampffmeyer, Robert Jenssen) introduced ProtoX-AD, a prototype-based self-explainable framework for self-supervised time series anomaly detection (TSAD). The key innovation is using learned prototypes — representative patterns from normal data — to explain why a time point is flagged as anomalous. Unlike existing black-box methods that only output anomaly scores, ProtoX-AD provides a set of prototypes that best match the anomalous region, giving users insight into the anomaly's characteristics (e.g., sudden spikes, flatlining, or frequency changes). The framework is transformation-aware, meaning it explicitly models how different data transformations (like scaling, warping, or noise addition) affect both detection performance and the interpretability of explanations.

ProtoX-AD achieves detection accuracy comparable to state-of-the-art black-box TSAD methods on synthetic and real-world datasets (e.g., NASA's MSL, Yahoo S5, and SWaT). Crucially, its explanations are more consistent and semantically meaningful than those from existing explainable baselines like gradient-based or attention-based methods. The paper (26 pages, 8 figures) includes ablation studies showing how prototype count and transformation design trade off between performance and explainability. Code is publicly available on GitHub, enabling immediate experimentation. This work addresses a critical gap in high-stakes domains like manufacturing, finance, or healthcare, where understanding why an anomaly occurred is as important as detecting it.