Abstract: Traditional hypergraph neural network is difficult to extract node features with high degree of correlation outside the direct neighborhood of nodes, which leads to incomplete global feature information. Dynamic hypergraph neural network (DHGNN) is improved, and a dynamic hypergraph neural network with joint graph random walk and skip connection (RWS-DHGNN) is proposed. RWS-DHGNN was used to classify non Euclidean data. Based on the DHGNN, graph random walk was introduced into the network to effectively obtain the node features with high degree of correlation outside the direct neighborhood of nodes. Meanwhile, the idea of residual network was introduced, and the residual structure was formed by adding skip connections at the vertex convolution of hypergraph. RWS-DHGNN gave full play to the advantages of graph structure and hypergraph structure. RWS-DHGNN was compared with GCN, HGNN, GAT and DHGNN on Cora dataset. The experimental results show that RWS-DHGNN can effectively improve the classification accuracy.