Research On Hypergraph Neural Networks Based On Page Rank Propagation Mechanism

In recent years,graph neural network has been widely concerned,and great achievements have been made in graph representation learning,However,with the continuous expansion of graph neural network application,the traditional graph structure defined point and point pair has unable to meet the high-order complex relationship between description data,so a new construction method needs to be introduced to redefine the connection between points.In contrast to traditional graphs,hypergraphs define pairwise point-to-point relationships as complex connections,and hyperedges can connect any associated nodes,presenting patterns of multiple nodes’ association.Compared with ordinary graphs,hypergraphs can accurately describe the relationships between multivariate-associated objects,retaining more effective information when describing higher-order data correlations.Hypergraph neural networks can be used to process complex data,capturing correlations between higher-order data using hypergraph structures.However,the current model is still constructed based on a shallow convolutional neural network,and each node can only have a small part but not gather neighbor node information in a wide range.In order to solve the above problems,based on the hypergraph neural network,this paper combines Hypergraph Neural Networks with Page Rank propagation mechanism,and constructs a Hypergraph Neural network based on Page Rank propagation mechanism by using personalized improved propagation scheme.The designed model maintains an effective attention to the root node while expanding the learning neighborhood,has a very large and tunable range of neighbors,and avoids the over-smoothing problem.In addition,the propagation process is separated from the network construction,so that a very wide propagation range can be achieved without improving the network itself.To verify the model validity,this paper carried out classification experiments on four public datasets.Experimental results show that our algorithm achieved the classification accuracy of 93.07%,85.79%,83.1% and 81.7%,up to 0.47%,8.59%,1.5% and 1.6% compared with HGNN method,respectively,and overcame the over-smoothing problem.In parameter analysis experiment,HGNNP tends to be stable with the increase of network layers.