Graph Attention Contrastive Learning for Few-Shot Semi-Supervised Node Classification

While graph neural networks (GNNs) excel in graph data analysis, most methods rely on extensive labeled data, which is often impractical due to labeling costs and complexity. To address this, we propose DynaMetaGAT, a novel framework integrating meta-learning to overcome overfitting and feature smoothing in low-label scenarios. DynaMetaGAT leverages a meta-learning-guided label propagation mechanism to generate high-precision pseudo-labels, expanding limited labeled data while ensuring broad feature perception. The model combines a multi-head attention GAT for dynamic neighbor feature clustering and a contrastive learning module to enhance node discriminability through positive-negative sample pairs. These components, optimized via two-layer meta-learning and adaptive label propagation, significantly improve featurelabel mapping efficiency. Experiments on benchmark datasets demonstrate DynaMetaGAT's superior performance, offering a robust solution for semi-supervised learning with scarce labels.