2026-09-10T09:45:01Z https://drops.dagstuhl.de/oai

oai:drops-oai.dagstuhl.de:27575 2026-09-10T05:38:43Z ddc:004 open_access

Spatiotemporal Adaptive Learning for Flow-Based Graph Neural Networks (Short Paper) Li, Fengjiao Wu, Meiliu Basiri, Ana spatiotemporal adaptive learning graph neural networks multi-graph representation GeoAI flow-based networks Spatial systems can often be represented in more than one relational way. For the same urban system, geographic adjacency, road connectivity, and mobility interaction provide different relational representations of how places are connected. However, most graph-based models either assume a single graph or combine multiple graphs without distinguishing their relative importance. This paper treats multi-graph spatiotemporal modelling as a problem of spatial perspective selection. We introduce adaptive spatial weights (α) to learn the relative importance of predefined relational graphs, and temporal weights (β) in the loss function to distribute emphasis across forecasting horizons. Using COVID-19 dynamics in Glasgow and Edinburgh as a case study, we find that learned spatial weighting patterns differ across cities. Glasgow shows a more balanced configuration, whereas Edinburgh places stronger emphasis on road connectivity. Compared with single-graph and equal-weight baselines, the adaptive model achieves comparable performance with improved stability while making graph weighting explicit. These results suggest that multi-graph weighting is a context-dependent process of selecting among alternative relational representations. Schloss Dagstuhl – Leibniz-Zentrum für Informatik Fengjiao Li and Meiliu Wu and Ana Basiri 2026 Is Part Of LIPIcs, Volume 393, 17th International Conference on Spatial Information Theory (COSIT 2026) InProceedings Text doc-type:ResearchArticle publishedVersion application/pdf doi:10.4230/LIPIcs.COSIT.2026.31 urn:nbn:de:0030-drops-275759 https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.COSIT.2026.31 eng https://creativecommons.org/licenses/by/4.0/legalcode