Totally Dynamic Hypergraph Neural Networks

As an extension of graphs, hypergraphs can naturally represent multi-relationships and have great application prospects in real life. The static hypergraph neural network relies too much on the initialized hypergraph structure and cannot mine hidden relationships within the data; the dynamic hypergraph neural network optimizes the hypergraph structure in the process of model iteration and can mine more information. However, the existing dynamic hypergraph neural networks ignore the features of hyperedges and cannot adjust the number of hyperedges, which proposes limitations when adjusting hypergraphs. We propose a novel hypergraph neural network that can adjust the number of hyperedges while optimizing the hypergraph structure. Our method focuses on hyperedge features and learns their feature distribution rather than fixed hyperedge features. The hyperedge is obtained by sampling from the learned distribution, and then the hypergraph is constructed according to the attention coefficient of sampled hyperedges and nodes, and finally, the node features are updated using the hypergraph convolution algorithm. Experimental results demonstrate the effectiveness of our method.