How Geospatial Effects Can Affect Predictive Learning: A Novel Predictability Framework Consisting of Geographic Entropy and Knowledge Graph

Spatiotemporal prediction (STP) is a fundamental research topic in GIScience, providing estimations of unobserved phenomena across space and time. Geospatial effects—spatial dependence, heterogeneity, and geographical similarity—underpin established spatially explicit learning frameworks, namely spatial dependence learning, regional learning, and transfer learning. A theoretical framework is lacking, however, to quantify these effects’ contributions to the learning process and to guide their optimal utilization in a spatially explicit learning framework construction. To fill this gap, we develop a novel predictability framework that integrates STP learning with geospatial effects. In this framework, we propose geographic entropy methods to quantify data predictability under geospatial effects. Using these predictability values, we construct predictability structures (autocorrelation, stratification, and teleconnection). Then, a knowledge graph and its corresponding knowledge services are established from these predictability structures. The knowledge services provide predictability knowledge to optimize spatially explicit learning frameworks, thereby enhancing STP performance. Experiments on a human-activity data set show that our framework significantly improves spatially explicit learning frameworks, achieving average accuracy gains of 4.40 percent for dependence learning, 20.99 percent for regional learning, and 18.21 percent for transfer learning.