Heterogeneous hypernetwork representation learning with feature awareness

Most existing methods for heterogeneous hypernetwork representation learning overlook the intricate correlations between the hyperedges and among the constituent nodes, where the hyperedges represent explicit higher-order tuple relationships. To address this limitation, we propose HRFA, a novel heterogeneous hypernetwork representation learning method with feature awareness. HRFA captures these critical correlations from dual perspectives to generate node representation vectors that are rich in both structural and node-specific features. Specifically, our approach consists of three core components. First, a structural feature-aware model is designed to capture the correlations between the hyperedges, learning hyperedge representation vectors that are subsequently transformed into node representation vectors rich in structural features. Second, a node feature-aware model is designed to capture the correlations among the nodes associated with the same hyperedges, yielding node representation vectors rich in intrinsic node attributes. Finally, a multi-feature fusion model integrates the node representation vectors from the above two models, and further calculates the intra-group tightness of these higher-order tuples to refine and update the final node representation vectors. Extensive experiments on four real-world hypernetwork datasets validate the effectiveness of HRFA. For the link prediction task, HRFA surpasses the optimal baseline methods with performance gains of 2.71%, 0.24%, and 2.63% on the MovieLens, drug, and wordnet datasets respectively. For the hypernetwork reconstruction task, HRFA outperforms most baselines, with significant improvements over the optimal baseline method under certain reconstruction ratios on the GPS and drug datasets.