TY - JOUR
T1 - Homophily modulates double descent generalization in graph convolution networks
AU - Shi, Cheng
AU - Pan, Liming
AU - Hu, Hong
AU - Dokmanić, Ivan
N1 - Publisher Copyright:
Copyright © 2024 the Author(s).
PY - 2024
Y1 - 2024
N2 - 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.
AB - 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.
KW - double descent
KW - graph neural network
KW - homophily
KW - statistical mechanics
KW - stochastic block model
UR - http://www.scopus.com/inward/record.url?scp=85185236694&partnerID=8YFLogxK
U2 - 10.1073/pnas.2309504121
DO - 10.1073/pnas.2309504121
M3 - Article
C2 - 38346190
AN - SCOPUS:85185236694
SN - 0027-8424
VL - 121
JO - Proceedings of the National Academy of Sciences of the United States of America
JF - Proceedings of the National Academy of Sciences of the United States of America
IS - 8
M1 - e2309504121
ER -