EHGFL: Contrastive distillation for efficient heterogeneous graph few-shot learning

Heterogeneous graphs (HGs), as a general modeling paradigm for multi-typed entities and complex interactions in ubiquitous real-world networks, have attracted extensive research enthusiasm. While self-supervised learning (SSL) on heterogeneous graph neural networks (HGNNs) demonstrates promising performance, existing SSL approaches face critical limitations: (1) The inherent multiple relation types and meta-path aggregations in HGNNs create prohibitive training and inference costs that restrict scalability; (2) The local message-passing paradigm confines information propagation to immediate neighborhoods, limiting the model’s ability to capture long-range dependencies and global structural patterns essential for complex reasoning tasks; (3) Task specialization necessitates costly fine-tuning of the HGNN backbones. To address these issues, we propose Efficient Heterogeneous Graph Few-shot Learning (EHGFL) to improve HGNNs’ scalability and global-structure modeling capabilities. Specifically, EHGFL first employs instance discrimination contrastive learning for self-supervised pretraining of HGNNs. To enhance efficiency, we introduce a novel cross-model contrastive distillation mechanism that transfers HGNNs’ heterogeneous structure modeling ability to a concise, globally-structure-aware multilayer perceptron. This feature-space distillation process preserves heterogeneous structure representations while avoiding expensive neighborhood aggregation and enhancing global feature awareness. Furthermore, to bridge the gap between pretraining objectives and downstream tasks, we adopt prompt tuning techniques specifically designed for the student model, enabling effective adaptation with limited labeled data. Extensive experiments on two real-world HG datasets demonstrate that the proposed EHGFL framework substantially accelerates training and inference while achieving superior few-shot node classification accuracy compared to state-of-the-art baselines.