Graph contrastive learning with virtual nodes for few-shot semi-supervised classification.

Abstract

Contrastive learning has shown strong potential in graph few-shot semi-supervised learning (GFSSL), due to its ability to leverage limited labeled data. However, its performance is often hindered by two critical issues: insufficient supervisory signals and the prevalence of imbalanced contrastive pairs, particularly the scarcity of positive samples. These can lead to representation collapse, where anchor nodes are aligned with only specific subclass features. To overcome these limitations, we propose a novel GFSSL framework that introduces label-guided virtual nodes within a supervised contrastive learning paradigm. Specifically, we first pre-train a graph encoder to mine high-confidence pseudo-labels from local neighborhoods, thereby supplementing the scarce supervision with reliable pseudo-labels. Then, we introduce label-guided virtual nodes, which act as class-level proxies that aggregate features across different subclasses. These virtual nodes are integrated into the graph to connect semantically similar nodes, facilitating cross-subclass feature propagation. Furthermore, we expand the number of positive samples by incorporating virtual nodes as positive samples in contrastive learning to provide anchor nodes with richer features from other subclasses, effectively preventing nodes from collapsing into a single subclass. Comprehensive experiments demonstrate that our model outperforms the best baseline method by an average of 2.06% in six datasets, highlighting the effectiveness of our approach.