Local Co-location Pattern Mining Based on Regional Embedding

Local co-location pattern (LCP) presents the spatial correlation between various categories in local regions. Regional partitioning is a pivotal step in LCP mining. Existing regional partitioning methods may ignore potential LCPs due to subjective elements. Additionally, with the diversity of geographic data increases, previous mining techniques disregarded the semantic information within the data, and limited the interpretability of local regions and LCPs. In response to these issues, this paper introduces an approach for LCP mining based on regional embedding. Initially, the entire study region is finely divided into local regions through natural data like road networks. Next, leveraging regional embedding techniques, local regions are embedded using human trajectory events, resulting in the creation of regional embedding vectors. Subsequently, the k-means method is employed to find functional clusters of local regions, and self-attention mechanisms is used for functional annotation. Then, the semantic LCPs aremined in these annotated local regions. Experiments on real-world datasets comprising urban population trajectories and Points of Interest (POI) confirm the efficiency and interpretability of the proposed framework for LCP mining based on regional embedding.