Dual-Kernel Graph Community Contrastive Learning

The recent introduction of an efficient Graph Contrastive Learning (GCL) framework marks a significant advancement in the field of Graph Neural Networks (GNNs). Traditional GCL methods struggle with scalability due to their reliance on intensive message passing and the quadratic complexity of contrastive loss calculations. The proposed framework mitigates these issues by transforming the input graph into a compact network of interconnected node sets, preserving structural information across communities. By implementing a kernelized graph community contrastive loss with linear complexity, the framework facilitates effective information transfer among node sets, capturing hierarchical structural information. Additionally, the incorporation of knowledge distillation into the decoupled GNN architecture enhances inference speed while maintaining strong generalization performance. Extensive experiments conducted on sixteen real-world datasets demonstrate that this innovative method surpasses…