Abstract
We propose an unsupervised learning framework for online quickest Change Point Detection (qCPD) for multi-entity, multivariate time series data, a problem setting in which system-level behavioral shifts must be detected across a population of interacting entities whose number and behavior may be uncertain or evolve. This setting, motivated by applications such as crowd monitoring where traditional sensing methods (e.g., video surveillance) may be infeasible, is not addressed by existing single-stream CPD methods or multi-entity anomaly detection techniques. We introduce IDfN-to-SWAS, a two-stage detection pipeline in which entity-level deviations from nominal behavior, Individual Deviation from Normality (IDfN) scores, are aggregated into a System-Wide Anomaly Score (SWAS). A CUSUM-based sequential detector is applied to the SWAS to identify persistent or abrupt system-level changes with low detection delay. The framework requires only training data obtained under normal system operation to determine the model required for IDfN calculation. Importantly, it naturally supports variable entity counts without retraining and is compatible with a wide range of anomaly scoring backbones. Evaluations on synthetic datasets and crowd simulations, explicitly designed for anomaly detection in group behaviors, demonstrate that our method accurately detects system-level changes, offering a scalable and privacy-preserving solution for monitoring complex multi-entity systems. To support reproducible benchmarking in this emerging problem domain, we additionally release three first-of-their-kind multi-entity multivariate datasets explicitly designed for evaluating system-level CPD in interactive multi-entity systems.