CLEAR: Graph contrastive learning with electrostatic self-adaptive repulsion

doi:10.1016/j.eswa.2025.130267

CLEAR: Graph contrastive learning with electrostatic self-adaptive repulsion

2025-11-08

0

PRE

AI

DOI：10.1016/j.eswa.2025.130267

Original

Original request for help

Save

Abstract

En 中文

Graph contrastive learning (GCL) optimizes representation spaces by drawing together augmented views of the same node (alignment) while spreading representations across the space (uniformity). Prior work has formalized these properties and shown their importance for contrastive objectives; however, many GCL methods either address them implicitly via pairwise losses and sampling or rely on heuristic regularizers that can be hard to interpret and tune. We present CLEAR (Electrostatic Adaptive Repulsion), a physics-inspired framework that explicitly optimizes the alignment - uniformity trade-off on the unit hypersphere. CLEAR models uniformity through a Coulomb-like potential with learnable, data-dependent charges and a dynamic charge center, yielding an isotropic, rotation-invariant repulsive field with a numerically stable surrogate objective. In parallel, a charge-aware alignment module integrates node semantics and topology to adaptively strengthen intra-class consistency. Extensive experiments on five benchmarks show that CLEAR consistently improves node classification accuracy over strong baselines, while achieving favorable alignment and uniformity metrics and robust training dynamics.

CLEAR: Graph contrastive learning with electrostatic self-adaptive repulsion

Abstract

Journal

Organization