Towards adaptive information propagation and aggregation in hypergraph model for node classification

In recent years, hypergraph models have gained widespread attention in the hypergraph node classification task due to their ability to capture high-order node relationships. Nevertheless, most previous models are unaware of the potential pairwise node relationships in hypergraph data and fail to sufficiently mine such relationships during the information propagation process, leading to suboptimal performance. Moreover, the over-smoothing problem causes the loss of distinctive features of nodes during the information aggregation process and limits the ability of previous models to extract and leverage valuable information from high-hop neighbors, resulting in restricted expressivity and performance. To tackle these problems, we propose a novel adaptive hypergraph neural network (AdaHGNN), that can adaptively mine potential pairwise node relationships and efficiently extract and adaptively aggregate information from high-hop neighbors for accurate hypergraph node classification. Specifically, we propose a new dual-view information propagation mechanism (IPM) that consists of graph-view IPM and hypergraph-view IPM to adaptively capture high-order node relationships and mine potential pairwise relationships, and fuse them through gate mechanism. We then simplify the conventional hypergraph convolution into a decoupled hypergraph convolution to efficiently extract information from high-hop neighbors. Finally, we customize a new adaptive aggregation method to adaptively aggregate valuable information from high-hop neighbors for classification. Extensive experiments conducted on four citation datasets demonstrate the superiority of the proposed model in accuracy and efficiency, and confirm its applicability.