Core Concepts
Introducing inter-graph connections during the self-supervised learning process significantly improves the performance of graph representation learning by enhancing manifold separation.
Stats
GIP significantly improves classification accuracy on the IMDB-MULTI dataset from below 60% to over 90%.
GIP consistently enhances the performance of four different self-supervised learning frameworks: MVGRL, G-BT, GRACE, and BGRL across six datasets.
GIP achieves near-perfect classification performance on several datasets.