Learning dual context aware POI representations for geographic mapping

Driven by artificial intelligence technologies, geospatial representation learning has become a new trend to better understand urban systems. Points of Interest (POI), as the current mainstream data in urban studies, plays an important role in these methods to discover urban characteristics. Existing studies on POI representation learning focus on spatial and type information, but overlook heterogeneous semantic interaction between POIs as well as hierarchical associations among types. To tackle these two problems, we propose a novel approach, called POI Dual Context Aware Neural Network (DCA) for learning POI representations by jointly embedding both spatial context and type context. For the spatial context of POIs, we introduce a distance decay effect constrained graph attention network as an encoder of DCA, which takes the heterogeneous semantic interaction and spatial proximity of POIs into account. For the type context of POIs, we propose a type hierarchical aggregation neural network architecture for DCA, and design a type infomax optimization objective following contrastive learning mechanism. The superiority of DCA is demonstrated in three geographic mapping tasks, including urban function mapping, region popularity mapping, and housing price mapping. This study provides a new insight to mine deep information from POIs, contributing to a better understanding of urban systems. The source code is released at http://github.com/quan-qin/DCA.