Homophily modulates double descent generalization in graph convolution networks
This work addresses the lack of theoretical understanding of GNN generalization for researchers and practitioners, with incremental improvements in performance on specific datasets.
The paper tackled the problem of understanding generalization in graph neural networks (GNNs) by analyzing double descent phenomena, showing that risk is shaped by graph noise, feature noise, and training labels, and used insights to improve state-of-the-art GNN performance on heterophilic datasets.
Graph neural networks (GNNs) excel in modeling relational data such as biological, social, and transportation networks, but the underpinnings of their success are not well understood. Traditional complexity measures from statistical learning theory fail to account for observed phenomena like the double descent or the impact of relational semantics on generalization error. Motivated by experimental observations of ``transductive'' double descent in key networks and datasets, we use analytical tools from statistical physics and random matrix theory to precisely characterize generalization in simple graph convolution networks on the contextual stochastic block model. Our results illuminate the nuances of learning on homophilic versus heterophilic data and predict double descent whose existence in GNNs has been questioned by recent work. We show how risk is shaped by the interplay between the graph noise, feature noise, and the number of training labels. Our findings apply beyond stylized models, capturing qualitative trends in real-world GNNs and datasets. As a case in point, we use our analytic insights to improve performance of state-of-the-art graph convolution networks on heterophilic datasets.