PROPGCL: Unleashing the Power of Propagation in Graph Contrastive Learning

ICLR 2026 Conference Submission17154 Authors

Keywords: Graph Contrastive Learning

Abstract: Graph contrastive learning (GCL) has recently gained substantial attention, leading to the development of various methodologies. In this work, we reveal that a simple training-free propagation operator PROP achieves competitive results over dedicatedly designed GCL methods across diverse node classification benchmarks. We elucidate PROP’s effectiveness by drawing connections with established graph learning algorithms. By decoupling the propagation and transformation phases of graph neural networks, we find that the transformation weights are inadequately learned in GCL and perform no better than random. When the contrastive and downstream objects are misaligned, the attendance of transformation causes the overfitting to the contrastive loss and harms downstream performance. In light of these insights, we remove the transformation entirely and introduce an efficient GCL method termed PROPGCL. We provide theoretical guarantees for PROPGCL and demonstrate its effectiveness through a comprehensive evaluation of node classification tasks.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 17154