Learnable Coreset Selection for Graph Active Learning

Graph Neural Networks (GNNs) have demonstrated their effectiveness in a variety of graph-based tasks. However, their performance heavily depends on the availability of a sufficient amount of labeled data, which is often costly to acquire in real-world applications. To tackle this, GNN-based Active Learning (AL) methods aim to enhance labeling efficiency by selecting the most informative nodes for labeling. However, existing methods often rely on heuristic or implicit approaches that fail to fully capture the influence of labeled data on unlabeled nodes, thereby limiting their adaptability across diverse graph types. In this paper, we propose LearnAL, a Learnable coreset labeling framework for graph Active Learning to address these limitations. Unlike traditional heuristic-based methods, LearnAL explicitly models the correlations between labeled and unlabeled nodes using an attention architecture, linking these correlations directly to prediction performance. Leveraging global influence (attention) scores, LearnAL selects and labels samples that maximize representational diversity, enhancing sample coverage. We provide theoretical analysis demonstrating that this attention-based selection reduces the covering radius bound, improving prediction performance on unlabeled data. Our experimental results show that the labeled coreset significantly enhances the generalizability of various graph models across different graph datasets, as well as CNN models in image classification tasks.