Geographic Entropy and Knowledge Graph

Spatio-temporal prediction (STP) is a fundamental research area in Geographic Information Science (GIS), offering estimations of unobserved phenomena across space and time. STP models are extensively applied to address geographic challenges such as weather forecasting and hazard warnings. Despite accuracy gains from data-driven AI technologies, current STP research often neglects the contributions of geospatial effects, which have inspired the development of GIS-style STP models, including dependence learning from the spatial proximity effect, regional learning from the spatial heterogeneity effect, and transfer learning from the geographic similarity effect. This study seeks to determine whether these geospatial effects enhance STP performance and how they can be leveraged to optimize GIS-style model design. We develop a predictability framework using geographic entropy (GE) for the former question and a knowledge graph KG for the latter. Specifically, GE comprises three entropy methods to evaluate changes in predictability under the influence of three geospatial effects. These results are organized using KG to provide knowledge services that optimize the design of three GIS-style models. Experiments using a real-world dataset consisting of five human activities demonstrate the effectiveness of our framework. Specifically, our framework quantifies gains or reductions in predictability under geospatial effects and depicts their influence structure via KG.