[3Xin4] Poster session 1

[3Xin4-76] An Initial Evaluation on Self-Supervised Learning for Heterophily Graph

〇Shuichiro Haruta¹, Tatsuya Konishi¹, Kota Matsui^2,1, Mori Kurokawa¹ (1.KDDI Research, Inc., 2.Nagoya Univ.)

Keywords:GNN

In recent years, self-supervised learning methods in graph attract attention.
At the same time, it is known that general GNNs (Graph Neural Networks) cannot deal with heterophily graphs which do not share the similar attributes among neighbor nodes.
In this paper, we investigate whether two representative graph self-supervised learning methods, DGI and GRACE, can handle heterophily graphs from the perspective of the node classification task.
Through experiments on both synthetic and real-world datasets, we show that these methods are unable to handle heterophily graphs. We conclude there is a need for graph self-supervised learning methods that can address heterophily graphs.