SymphoNEI: Symphony of Node and Edge Inductive Representations on Large Heterophilic Graphs

Homophily-based GNNs do not adequately interpret heterophilic graphs, i.e., classes or attributes between adjacent nodes are often different. Therefore, GNNs for heterophilic graphs have recently been proposed, but several limitations remain. Applying GNNs to large-scale graphs is performed in inductive settings with several sampling methods. However, most heterophilic GNNs do not perform well when combined with sampling methods and do not generalize to unseen nodes because they only consider the transductive setting. In addition, heterophilic GNNs only extract information from nodes and graph structures. They do not represent edges, which could be essential elements in the graph. To challenge the issues, we propose a new model SymphoNEI. SymphoNEI successfully embeds large heterophilic graphs in the inductive setting by learning both node and edge representations. To the best of our knowledge, SymphoNEI is the first of the heterophilic GNNs to represent the edges and evaluate inductive node classification on large heterophilic graphs with and without edge features. We used eight large heterophilic graphs for model evaluation: five without edge features and three with edge features. Extensive experimental results indicate that SymphoNEI achieves state-of-the-art performance for inductive node classification on large heterophilic graphs.