Generative self-supervised learning in hypergraphs with adaptive masking strategies

Hypergraph Neural Networks (HGNNs) have been widely applied to analyze data structured as hypergraphs, excelling in handling complex node relationships. However, the scarcity of truth labels in small-sample scenarios is common, which limits the generalization ability of HGNNs. To tackle this problem, we introduce a new generative self-supervised learning framework, AHGL, designed to enhance model learning through rich masked inputs. AHGL introduces a masked feature recovery strategy that encourages the model to capture fine-grained high-order relationships in hypergraphs. Additionally, the proposed stage incremental masking technique and adaptive masking strategy effectively enhance the diversity and adaptability of the model's inputs. Results from the experiments indicate that AHGL delivers notable enhancements in performance across various datasets.