Climate-Aware Self-Retrospective Representation Learning for Spatio-Temporal Epidemic Forecasting

Spatio-temporal epidemic forecasting aims to predict future outbreak trajectories across interconnected regions from historical epidemiological observations and meteorological covariates. However, existing approaches often fail to preserve historically salient epidemic states or to fully exploit delayed and region-varying meteorological associations, leading to unstable temporal representations and insufficient meteorological-context-aware spatio-temporal context for prediction at later forecast horizons. In this paper, we propose CASRL, a Climate-Aware Self-Retrospective Representation Learning network for stable and meteorological-context-aware spatio-temporal epidemic forecasting. CASRL first employs a Self-Retrospective Epidemic Encoder (SREE) to retrospectively aggregate historically salient epidemic states through query-guided weighting and adaptive gating, thereby preserving informative historical epidemic states within the look-back window. It then introduces a Climate-Adaptive Graph Message Passing (CAGMP) module that breaks away from traditional passive feature concatenation. Instead, it constructs a separate meteorological-view predictive graph conditioned on the static spatial prior and adaptively fuses it with the incidence-associated topology to model complex cross-regional predictive associations. By integrating self-retrospective epidemic representations with meteorological-view spatio-temporal interactions, CASRL produces forecasts with improved predictive stability at later forecast horizons. Extensive experiments on two public influenza benchmarks show that CASRL is competitive at shorter forecast horizons and provides clearer advantages at later forecast horizons, particularly in phase-alignment-related evaluation and 15-week-ahead forecasting.