A pseudo-labeling approach based on knowledge distillation for graph few-shot learning

Graph-based few-shot node classification (FSNC) has emerged as a promising solution to the challenge of limited labeled nodes in complex network analysis. Although existing pseudo-labeling FSNC methods have shown encouraging results, they often struggle with wrong or over-confident pseudo-labels, which can negatively impact model generalization. To overcome these limitations, we propose PLD-FSNC, a novel pseudo-labeling FSNC framework leveraging knowledge distillation. Our PLD-FSNC framework is composed of two modules, i.e., embedding transfer and pseudo-label improvement. The embedding transfer module transfers knowledge from a pre-trained source model to a target model, enhancing pseudo-label selection quality. The pseudo-label improvement module mitigates the impact of wrong and over-confident pseudo-labels by using soft labels from the source model to supervise the target model's predictions. We also provide theoretical justification for our pseudo-label improvement module and demonstrate its effectiveness through extensive experiments on six real-world datasets.