A framework for out-of-distribution-robust spatiotemporal prediction over graphs that combines two OOD principles in one model. Invariance existence is enforced by an invariant-risk-minimization penalty; environment diversity is supplied by a graph-based diffusion model acting as an environment augmentor, guided by a causal mask generator that separates causal (invariant) from environment (spurious) features. The invariance penalty computed on the diffusion-augmented data regularizes training of the spatiotemporal predictor, yielding interpretable, distribution-shift-robust forecasts validated on synthetic structural-causal-model data and human-mobility/traffic datasets.
A framework (diffIRM) for out-of-distribution-robust spatiotemporal prediction over graphs, e.g., traffic and human-mobility forecasting whose patterns shift over time. It unites two OOD principles: …