Gradient boosted graph convolutional network on heterophilic graph

Abstract

Graph Neural Networks (GNNs) are impressive models that have been highly successful in performing graphical analysis and learning. However, GNNs are known to be outstanding in learning from homophilic graphs but are subpar in learning from heterophilic graphs. On the other hand, Gradient Boosted Decision Trees (GBDTs) have become the best-performing model in dealing with heterogeneous tabular data. But will GBDTs retain that superiority when working with heterophilic graphs? This project proposes an alternative model to learn from heterophilic graphs, combining both GNNs and GBDTs. GBDTs will only focus on training the node features of the heterophilic graphs, passing the refined node features to the GNN to improve on the graph structure. After experimentation and comparison with GNN models, Graph Convolutional Network (GCN) in the case of this project, the proposed alternative model has shown a reasonable increase in performance in learning from heterophilic graphs.