IRGCL： Information Refinement Graph Contrastive Learning

Yilin Li, Qingshuai Feng, Zhen Tao, Ziyue Qiao, Qingqiang Sun

Keywords: feature selection, structure learning, degree aware, graph contrastive learning, single view

Abstract:

Graph contrastive learning (GCL) has emerged as a leading paradigm in unsuper-
vised graph representation learning (UGRL), yet existing contrastive approaches
remain vulnerable to three persistent challenges: noisy features that distort simi-
larity measures, unreliable structures that contain spurious edges, and degree im-
balance that biases representation quality. We propose Information-Refinement
Graph Contrastive Learning (IRGCL), a single-view contrastive learning frame-
work that simultaneously addresses these challenges and effectively generalizes
across key graph learning tasks, including node classification, clustering, and link
prediction. IRGCL integrates three complementary components: (i) structure-
consistent feature selection to filter out redundant or noisy attributes; (ii) high-
confidence structure learning to refine graph neighborhoods; and (iii) degree-
aware focal contrastive learning to balance learning across low- and high-degree
nodes. Extensive experiments on diverse benchmarks demonstrate that IRGCL
consistently outperforms state-of-the-art baselines, and ablation studies confirm
the distinct and complementary benefits of each component, highlighting the ne-
cessity of jointly addressing feature quality, structural reliability, and degree im-
balance. Code is available at https://anonymous.4open.science/r/IRGCL-01F8.

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

Submission Number: 6792