Enhanced sample selection in graph contrastive learning with attribute-structure fusion

Graph contrastive learning (GCL) has emerged as a crucial framework for advancing self-supervised graph representation learning. However, a critical challenge in GCL stems from insufficient positive sample selection, which results in limited supervision signals and significantly hinders the representation learning process. To address this issue, we propose the Attribute-Structure Fusion Sampling (ASFS) module - a novel, plug-and-play framework designed for seamless integration with existing GCL models. Specifically, it leverages attribute and structural information to comprehensively select high-quality positive samples for each node. To further enhance the reliability of positive samples, we employ a dynamic self-correction mechanism based on deep clustering, which iteratively refines the selected samples during the training process. Extensive experiments on six benchmark graph datasets demonstrate the effectiveness and adaptability of ASFS in addressing the challenge of positive sample selection in GCL.