Graph Transformation Augmentation for Contrastive Learning of Graph-Level Representation: An Initial Exploration

Abstract

Contrastive learning on the image data becomes a representative method of self-supervised learning to pre-train a neural encoder from data or model perspective(s). However, the data-perspective method in the graph domain is less explored because graph data augmentation is not as mature as image data augmentation. In this paper, we propose a transformation-based graph data augmentation, which is named Graph Transformation Augmentation (GTA). GTA will preserve the information of the graph spectrum instead of the subgraph information. GTA has two types: Permutation Augmentation and Orthonormal Augmentation. Finally, we experimentally validate the workability of GTA on self-supervised representation learning, and GTA unintuitively preserves the graph semantics.