Representing Spatial Data with Graph Contrastive Learning

Large-scale geospatial data pave the way for machine learning algorithms, and a good representation is related to whether model effective. Hence, it critical task learn effective feature data. In this paper, we construct spatial graph from locations propose contrastive method location representations. Firstly, skeleton in order preserve primary structure of solve positioning bias problem remote sensing. Then, define novel mixed node centrality measure four augmentation methods based on measure. Finally, heterogeneous attention network aggregate information both structural neighborhood semantic separately. Extensive experiments datasets non-geospatial are conducted illustrate that proposed outperforms state-of-the-art baselines.