联合图随机游走和跳跃连接的动态超图神经网络

DYNAMIC HYPERGRAPH NEURAL NETWORK WITH JOINT GRAPH RANDOM WALK AND SKIP CONNECTION

针对传统超图神经网络难以提取节点直接邻域外关联度高的节点特征,导致全局特征信息不完整的问题,对动态超图神经网络(DHGNN)进行改进,提出联合图随机游走和跳跃连接的动态超图神经网络(RWS-DHGNN),用于非欧几里得数据的分类。该网络在DHGNN的基础上,引入了图随机游走,从而有效地获取直接邻域外关联度高的节点特征。同时,引入残差网络的思想在超图的顶点卷积处增加跳跃连接构成残差结构。所提网络模型充分发挥图结构和超图结构的优势。在Cora数据集的标准分割和随机分割上将所提网络与GCN、HGNN、GAT和DHGNN进行对比实验,实验结果表明,该模型可以有效提高分类准确率。

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.