Research On Link Prediction Of Heterogeneous Information Network Based On Graph Attention Mechanism

Link prediction is one of the important tasks in the field of heterogeneous graph data analysis,and can be widely used in many practical scenarios.At present,many studies on the link prediction model of heterogeneous information networks have not considered the unequal impact of different types of nodes and edges on the target node,and have not considered the evolution of semantic features and topological structure on the target node in the time dimension.The impact will cause the low accuracy of the link prediction effect and the high time complexity.In response to the above problems,this paper,through research,proposes a heterogeneous information network link prediction model based on graph attention mechanism,and improves and optimizes the model.At the same time,it proposes a link prediction model for dynamic heterogeneous information networks.Solve the problem of how to select key features for aggregation when the target node has a large number of neighbors in a heterogeneous information network.The main work of this paper is as follows:First,a heterogeneous information network link prediction model based on graph attention mechanism is proposed.The model uses the graph attention mechanism to perform weighted distribution through the intra-meta-path aggregation and the intermeta-path aggregation of the target node,and the vector generated by the meta-path adjacency matrix strengthens the information of the network topology structure of the target node's feature embedding.The link prediction effect of the model was verified by using public data sets and data sets in the open source community field.Second,an improved optimization model based on adaptive heterogeneous information is proposed.In view of the adverse effects of the manual selection of metapaths on the embedding of target nodes,and the ignoring of multiple different types of links between the same node,the model modifies the calculation target of the attention mechanism,that is,the first-order meta-path to the target node Neighbors calculate the multi-head attention matrix according to different node types and edge types,and adaptively select important neighbor nodes and edges for feature aggregation.Compared with the basic model proposed in the previous article and the work of other scholars,the prediction effect index of the optimized model has increased by 1.86%-5.37%.Third,a link prediction model for dynamic heterogeneous information networks is proposed.The model considers the semantic features of nodes and the evolution of network topology features from the time dimension,and aggregates target nodes in the three aspects of neighborhood,topology,and time dimension through graph convolutional layer,long-term short-term memory layer,and time-series attention layer.Feature embedding.Experiments show that compared with the basic model and other models based on dynamic heterogeneous information network,the prediction effect index of this model is increased by 4.28% and 0.88%-1.28% respectively.The above three main works use the attention mechanism to weight and assign different influencing factors in the process of target node feature aggregation from the meta-path,node and edge categories,and time dimension.The experiment proves that the related model proposed in this paper has good performance in the link prediction task of heterogeneous information network.