SCALE: Style-Causal Disentanglement with Adaptive Lifelong Expert for Online Latent-domain Anomaly Detection

Time-series anomaly detection in real-world streams is often challenged by evolving operating conditions, where distribution shifts can be easily mistaken for anomalies. Due to this, we study a new problem, online latent-domain anomaly detection, where domain labels and shift times are unobserved, the number of domains may grow over time, and the detector must adapt and detect anomaly in a single-pass stream. To address this setting, Style–Causal Disentanglement with Adaptive Lifelong Expert (SCALE) is introduced as an online framework that disentangles each window into domain-invariant causal features and domain-specific style features, supports continual adaptation with a dynamic expandable style-expert pool, and separates domain drift from true anomalies using two decoupled criteria: Sparse Likelihood Regret (SLR) for drift identification and expert routing, and Counterfactual Prediction Residual (CPR) for final anomaly scoring. Theoretical properties are established to show that SLR reliably detects domain drift and, when combined with CPR, further distinguishes domain shift from true anomalies. Extensive experiments are conducted on diverse real-world datasets, together with a new multi-domain time-series anomaly detection benchmark for systematically evaluating online latent-domain scenarios.