[3A1-GS-10] AI application:

[3A1-GS-10-05] Towards Spatial Data Anaysis with Equivariant GNN

〇Taishi Yamamoto¹, Yohei Kodama², Yuki Akeyama², Kyohei Atarashi¹, Hisashi Kashima¹, Koh Takeuchi¹ (1. Kyoto University, 2. West Japan Railway Company)

Keywords:GNN, Spatial Data

In the field of spatial data analysis, research has been conducted to make predictions from unstructured data with complex spatial dependencies. Recently, deep learning methods, particularly Graph Neural Networks (GNNs), have been widely utilized to extract complex and spatially dependent patterns. In this study, we address the problem of making predictions on general heterogeneous graphs whose nodes and edges have multiple types, and edges have features. An example of task with such heterogeneous graphs is predicting demand from a specific region to a specific station. For analyzing heterogeneous graphs, we propose a model that can effectively capture information about similar spatial relationships observed in different regions by extracting features equivariant to transformations such as rotation and translation. Experiments on datasets with strong spatial dependencies demonstrate that the proposed method outperforms existing methods in predictive performance.