Unveiling the Dynamic Interactions between Spatial Objects: A Graph Learning Approach with Evolving Constraints

Abstract

Discovering time-aware interactions among spatial objects is essential for various urban applications, such as offline advertising and public transport planning. Although previous studies have focused on identifying static relationships among spatial objects, little attention has been given to investigating dynamic location interactions. However, the availability of urban data through human activity creates new opportunities to understand the evolving relationships between connected objects. Therefore, we introduce a new problem of determining multiple interactions among spatial objects to address the challenge of integrating dynamic and spatial impact under interacting sparsity constraints. To tackle this problem, we propose a graph learning solution that leverages an Evolving Graph Neural Network (EGNN) consisting of two collaborative components: a Cross Spatial-Interaction Propagation (CSIP) and an Evolving Self-Supervised Learning (ESSL) module. CSIP enables aggregation within- and propagation across time segments to capture evolving context within a spatial scope from the perspective of message passing between placements. ESSL employs time-aware learning with global and local loss reduction and introduces an additional evolving constraint to consider sparsity of interactions in spatial representation learning. Experiments on two real-world datasets demonstrate the superiority of our approach over several state-of-the-art methods.