Mind the Gap: Mitigating the Distribution Gap in Graph Few-Shot Learning

Abstract

Prevailing supervised deep graph learning models often suffer from the issue of label scarcity, leading to performance degradation in the face of limited annotated data. Although numerous graph few-shot learning (GFL) methods have been developed to mitigate this problem, they tend to rely excessively on labeled data. This over-reliance on labeled data can result in impaired generalization ability in the test phase due to the existence of a distribution gap. Moreover, existing GFL methods lack a general purpose as their designs are coupled with task or data-specific characteristics. To address these shortcomings, we propose a novel Self-Distilled Graph Few-shot Learning framework (SDGFL) that is both general and effective. SDGFL leverages a self-distilled contrastive learning procedure to boost GFL. Specifically, our model first pre-trains a graph encoder with contrastive learning using unlabeled data. Later, the trained encoder is frozen as a teacher model to distill a student model with a contrastive loss. The distilled model is then fed to GFL. By learning data representation in a self-supervised manner, SDGFL effectively mitigates the distribution gap and enhances generalization ability. Furthermore, our proposed framework is task and data-independent, making it a versatile tool for general graph mining purposes. To evaluate the effectiveness of our proposed framework, we introduce an information-based measurement that quantifies its capability. Through comprehensive experiments, we demonstrate that SDGFL outperforms state-of-the-art baselines on various graph mining tasks across multiple datasets in the few-shot scenario. We also provide a quantitative measurement of SDGFL’s superior performance in comparison to existing methods.