Graph Neural Networks with Extreme Nodes Discrimination

Graph neural networks (GNNs) are a successful example of leveraging the underlying structure between samples to perform efficient semi-supervised learning. Though the spatial correlation of the nodes is inherently taken into account by the models’ architecture, structural correlations and their effects in learning remain a relatively overlooked topic. In this work, we propose a new approach to train a GNN, by separating the samples based on their structural importance, meaning discriminating for samples that belong in a higher tier in terms of a network centrality metric. Our proposed method is supported by recent theoretical findings based on Extreme Value Theory, that buttress the separation of extreme and regular samples in binary classification. Essentially we split a GNN into two parts, each trained and validated separately using extreme and regular nodes from the observed set. We perform experiments in the three most prevalent GNN models, using three well-known benchmark datasets and compare the predictions of the model with and without sample discrimination. The classification of extreme nodes is clearly benefited, validating the relevant theory. In contrast, the regular nodes are undermined, despite their significantly larger train set. Exploratory findings suggest the limited structure contained in regular samples to be a potential reason for this.