Integrating metro passenger flow data to improve the classification of urban functional regions using a heterogeneous graph neural network

The accurate classification of urban functional regions is pivotal for city planning and resource allocation. While previous studies have integrated multi-source data for classification, few have considered both the spatial proximity of functional regions and their functional relationships reflected by crowd mobility. To address this gap, we propose a heterogeneous graph neural network (HGNN) approach that integrates metro passenger flow data with points of interest (POIs) and buildings. Initially, we constructed a heterogeneous graph with blocks and metro stations as nodes, connected based on their spatial proximity and crowd flow interactions. Subsequently, the building morphological and POI categorical features and crowd mobility features were used as the descriptive features of the block and metro station nodes, respectively. Finally, an HGNN model is designed to process the graph and classify the block functional types using a semi-supervised learning method. Experiments on the datasets of Beijing City indicate that our proposed approach achieved an accuracy of 81.56%. Ablation experiments reveal that integrating metro passenger flow data improved the classification accuracy by 6.87%. These results demonstrate the necessity of integrating crowd flow data to infer region functions and the superiority of HGNN in fusing heterogeneous source data while considering their inter relationships.