Improved graph neural networks for spatial networks using structure-aware sampling

Abstract

Graph Neural Networks (GNNs) have received wide acclaim in recent times due to their performance on inference tasks for unstructured data. Typically, GNNs operate by exploiting local structural information in graphs and disregarding their global structure. This is influenced by assumptions of homophily and unbiased class distributions. As a result, this could impede model performance on noisy real-world graphs such as spatial graphs where these assumptions may not be sufficiently held. In this article, we study the problem of graph learning on spatial graphs. Particularly, we focus on transductive learning methods for the imbalanced case. Given the nature of these graphs, we hypothesize that taking the global structure of the graph into account when aggregating local information would be beneficial especially with respect to generalisability. Thus, we propose a novel approach to training GNNs for these type of graphs. We achieve this through a sampling technique: Structure-Aware Sampling (SAS), which leverages the intra-class and global-geodesic distances between nodes. We model the problem as a node classification one for street networks with high variance between class sizes. We evaluate our approach using large real-world graphs against state-of-the-art methods. In the majority of cases, our approach outperforms traditional methods by up to a mean F1-score of 20%.