Joint Urban Modeling With Graph Convolutional Networks and Crowdsourced Data: A Novel Approach

Graph Convolutional Networks (GCN) are a potent and adaptable tool for effectively processing and analyzing continuous spatial data. Despite the substantial potential of GCN in various domains, most existing spatial data prediction models are confined to defining weights solely based on distance. To overcome this limitation, this study proposes a novel approach to obtain the second-level embedding of Points of Interests (POIs) by employing Delaunay Triangulation (DT), Random Walk, and Skip-Gram model training. Subsequently, enhanced features are obtained through various aggregation strategies for regional embedding. The integrated grid data, including longitude and latitude coordinates, enhanced features, and target values, are then integrated. Finally, the GCN is utilized for training and fitting to achieve the final prediction target value. By considering the influence of weights on data prediction, this approach can more accurately reflect the distribution and relationships of data in the actual environment. Furthermore, we have experimentally validated the effectiveness of this approach, demonstrating that it significantly enhances the accuracy of spatial data prediction when compared to the original GCN model's approach.