Improving augmentation consistency for graph contrastive learning

Graph contrastive learning (GCL) enhances unsupervised graph representation by generating different contrastive views, in which properties of augmented nodes are required to be aligned with their anchors. However, we find that in some existing GCL methods, it is hard to inherit semantic and structural properties of graphs from anchor views due to inconsistent augmentation schemes, which may hurt node consistency in augmented views. In this paper, we present ConGCL to improve node consistency and enhance node classification. Specifically, we first consider context entailment, which integrates the semantic and structural properties to better mine the underlying consistency relationships of nodes. Beneficial from this, we then design a novel consistency improvement loss to maintain augmentation consistency agreement among positive node pairs under stochastic augmentation schemes. To investigate the effectiveness of ConGCL on improving augmentation consistency and enhancing node classification, we conduct empirical study and extensive experiments on benchmark datasets. The code is available at: https://github.com/brysonwx/ConGCL.