We are developing geospatiotemporal predictive models

Spatiotemporal graph neural networks (STGNNs) have been adopted for predictive analysis in various scientific domains. Despite their promising performance, the dominance of big (geo)spatiotemporal data with large and heterogeneous dimensions raise computational challenges to effective adoption, generalization and fine-tuning of graph models. Moreover, continuous and high quality historical data may not always exist for such generalization. This paper proposes a framework that can co-evolve historical geospatial temporal datasets and an STGNN model by (1) incorporating elevation features optimized for water systems, and (2) integrating and interacting geospatial data discovery and graph learning with a "rehearsal" mechanism, that automatically generalize STGNNs to broader areas. The process divides spatiotemporal data into regional fragments with inferrable features, and iteratively (1) augment sparse training data in terms of feature similarity, (2) explore the augmented data by a trial "rehearsing" of the current model to decide a fraction of data to be adopted, over which a consistently good accuracy is observed, and (3) generalize STGNNs with promising regional data, ensured by rehearsal performance. This exploratory process hence learns to decide when and where to generalize STGNNs, for cost-effective generalization. Using real-world datasets, we experimentally verify the effectiveness and efficiency of our rehearsal framework.