Dynamic hypergraph neural networks based on key hyperedges.

• The original and dynamic topology are both introduced into hypergraph-oriented GNN. • The role of key hyperedges in GNN is revealed; the corresponding algorithm is built. • A new benchmark GNN model by covering the information of key hyperedges is provided. The technique of graph/network embedding can help computer to efficiently analyze and process the complex graph data via vector operations. Graph Neural Network, which aggregates the topological information of the neighbourhoods of each node in a graph to implement graph/network embedding, has attracted wide attention. With the explosive growth of information, large amounts of data need to be expressed in the form of hypergraphs. As a result, the hypergraph neural networks arise at the historic moment. However, most current work is based on static hypergraph structure, making it hard to effectively transmit information. To address this problem, D ynamic H ypergraph Neural Networks based on K ey H yperedges (DHKH) model is proposed in this paper. Considering that the graph structure data in the real world is not uniformly distributed both semantically and structurally, we define the key hyperedge as the subgraph composed of a small number of key nodes and related edges in a graph. The key hyperedge can capture the key high-order structure information, which is able to enhance global topology expression. With the supporting of hyperedge and key hyperedge, DHKH can aggregate the high-order information and key information. In our experiments, DHKH shows good performance on multiple datasets, especially on the SZ dataset and LOS dataset which have inherently some key structures. [ABSTRACT FROM AUTHOR]