摘要
Considering the extreme class imbalance in real-world graphs, increasing attention has been paid to Few- Shot Node Classification (FSNC). However, existing methods in traditional setting face two critical issues. Firstly, random task sampling without considering relationships can lead to a lack of structures within each task, unsuitable for graph data. Secondly, to compensate, they inefficiently aggregate information across entire graph during each task adaptation, which contradicts the intention of few-shot learning. Thus, we propose anew setting, Query-focused FSNC (Q-FSNC), which ensures structural relevance during task sampling and better suits graph data. More importantly, it is efficient while challenging due to information scarcity from the restriction of few-shot, i.e., task adaptation is restricted to few intra-task nodes only. To address this, we propose Hierarchical Global to Local calibration (HGL), the first work in this field to leverage tri-granular global information at subclass-, class-, and superclass levels. HGL incorporates following two novel components. Feature calibration enhances node features by integrating hierarchical feature distributions, effectively addressing class-specific information scarcity. Classifier calibration refines predictions with global class covariances and pseudo-labels, thereby alleviating between-class information scarcity with comprehensive global between-class information. Extensive experiments in both FSNC and Q-FSNC settings show that HGL significantly outperforms competing methods across 6 datasets.
