Multi-Augmentation Contrastive Learning as Multi-Objective Optimization for Graph Neural Networks

Abstract Recently self-supervised learning is gaining popularity for Graph Neural Networks (GNN) by leveraging unlabeled data. Augmentation plays a key role in self-supervision. While there common set of image augmentation methods that preserve labels general, graph do not guarantee consistent semantics and are usually domain dependent. Existing GNN models often handpick small techniques limit the performance model. In this paper, we propose to wide range tasks, rely on Pareto optimality select balance among these possibly conflicting augmented versions, called P areto G raph C ontrastive L earning ( PGCL ) framework. We show while random selection same leads slow convergence or even divergence, converges much faster with lower error rate. Extensive experiments multiple datasets different domains scales demonstrate superior comparable PGCL.