Ego-Aware Graph Neural Network

Abstract

In scientific exploration and daily life, inter-connectivity between entities is ubiquitous, such as species via the food webs, people among social networks, and products and customers in e-commerce. Graphs and networks are natural constructs to model such linkages, which can capture the relational information in addition to features provided by individual nodes. Graph neural networks (GNNs) are a powerful computation tool for reasoning about interconnected data and can help us make better use of the information imbued in data collectively. GNNs have found many successful applications, such as predicting drug side-effects, classifying diseases, and predicting the function of proteins. Most existing GNN methods are designed for datasets with moderate node features, where individual node features provide limited information, and assortative graph datasets, where nodes and their neighbours are more likely to be similar. However, for graphs with rich node features, or for the disassortative graphs, where nodes and their neighbours tend to be different, the message passing process in GNNs might be susceptible to inferences from the neighbourhood, leading to performance deterioration. To address this issue, we find that ego networks contain an extra layer of information to further distinguish different ego nodes. Based on that, we propose a generic GNN model that can better utilize structural information of nodes' proximity to extract informative messages and resist contradictory ones from their neighbourhood. We analyze how node features and graph structure can influence the performance of GNN models. Experimental results provide insight on how our method outperforms the baselines. The presented model paves the way for incorporating ego networks' structural information into the learned graph representations, which brings GNNs with better performance and higher robustness over different datasets.