GCL-UP: Graph Contrastive Learning via Uncertainty-Aware Perturbation for Few-Shot Node Classification

Abstract

Few-shot node classification is a fundamental challenge for Graph Neural Networks (GNNs), where models must generalize from a limited number of labeled nodes. Under such label scarcity, conventional studies have employed Graph Contrastive Learning with graph view augmentation to improve representation learning. However, most existing Graph Contrastive Learning methods rely on naïve random perturbations for augmentation, which often produce task-irrelevant and less informative graph views. This limitation is particularly problematic in few-shot node classification, where the model is highly sensitive to low-quality graph views. To address this, we propose GCL-UP (Graph Contrastive Learning via Uncertainty-aware Perturbation), a framework that employs uncertainty-aware graph perturbation for multiple view augmentation in node-level contrastive learning. The key insight of GCL-UP is that prediction uncertainty serves as a principled, task-aware signal for identifying the most informative regions to perturb. By selectively adding edges in reliable (low-uncertainty) regions while pruning those in ambiguous (high-uncertainty) regions, our method generates augmented graph views that are more informative and semantically consistent. We evaluate GCL-UP on three benchmark datasets (Cora, Citeseer, and Coauthor-CS) across 1–5 shot settings, demonstrating competitive performance against state-of-the-art GNN models. Our results show that uncertainty-aware graph perturbation yields more discriminative and robust representation learning in few-shot node classification.