Fair Contrastive Learning on Graphs

Node representation learning plays a critical role in learning over graphs. Specifically, the success of contrastive learning methods in unsupervised node representation learning has been demonstrated for various tasks, which has led to increase in attention towards the field. Despite the increasing popularity, fairness is widely under-explored in the area. Motivated by this, this study proposes novel fairness-aware graph augmentations based on adaptive feature masking and edge deletion, in order to mitigate the bias in graph contrastive learning. Different fairness notions on graphs are introduced in the study to guide the designs of the proposed adaptive augmentation schemes. Moreover, it is quantitatively shown that the proposed feature masking scheme can reduce the intrinsic bias. Experimental results on four real-world networks are presented to show that the introduced augmentation frameworks can improve group fairness measures together with comparable classification accuracy to state-of-the-art graph contrastive learning studies for node classification.