Interpretable Regime Trajectories via Generative Graph State-Space Models

Forecasting the behavior of real-world spatiotemporal systems often requires not only accurate predictions but also interpretable regime trajectories, i.e. discrete states that describe how dynamics change over time. However, existing approaches often entangle space and time, obscuring regime structure or trading interpretability for scale. We introduce ReGraSS, a unified framework that learns discrete, interpretable latent regimes from spatiotemporal data, represented as dynamic graphs, combining variational training with strictly time-ordered state-space inference. Predictions are produced by a mixture-of-experts modulated by the inferred regime probabilities, enforcing regime-specific specialization and supporting interpretability. Trained with self-supervised one-step prediction, the model learns in label-scarce settings and provides calibrated uncertainty by estimating a distribution over discrete regimes. ReGraSS matches or surpasses state-of-the-art spatiotemporal baselines in one-step forecasting. It shows the smallest error spike at regime changes and the fastest recovery thereafter, indicating regime-level interpretability and reliable trajectory tracking without compromising accuracy. We believe our interpretable, uncertainty-aware framework for regime-aware forecasting on dynamic graphs has direct application in healthcare, finance, and epidemiology.