A community-aware graph neural network applied to geographical location-based representation learning and clustering within GIS

Abstract

For many years, geographical location-based clustering has continuously become a hot research topic in both geographic information systems (GIS) and machine learning (ML)/deep learning (DL) domains due to its applications for various real-world problems. Spatial density-based clustering methods are generally utilized to detect clusters from geographical data. Traditional spatial clustering methods encounter considerable limitations, especially when addressing sparse coordinate data and substantial fluctuations in geographic cluster density. To overcome these limitations, in this paper, we introduce a novel community-aware graph learning for geographical point embedding, naming: CGL4PE. By extensively analyzing nearby graph-modelled geospatial locations, our proposed CGL4PE model can efficiently capture the complex/multi-view structural latent representations of geographical locations. To do this, in our model we apply the integration between graph attention network (GAT) and graph isomorphism network (GIN) in the form of multi-view graph representation learning approach. Then, combining with location-based graph construction strategy, our CGL4PE model can directly improve the clustering performance of conventional spatial/density-based clustering methods. Different from previous location-based clustering techniques, our CGL4PE model is designed to better preserve the community-aware structural representations of constructed location-based graph. These community-aware structural representations are then utilized to extract more meaningful cluster information from geographical datasets, under the integrated graph representation learning and spatial density-based clustering approaches. The extensive empirical studies within real-world geographical datasets present the effectiveness of our proposed CGL4PE model in this paper.