Research On Node Embedding Feature Representation Based On Graph Neural Network

In recent years,representation learning has received extensive attention and research.Among them,approaches based graph neural network have made remarkable achievements in various fields.This thesis will focus on the feature representation of nodes on heterogeneous graph,and the key lies in how to fully obtain heterogeneous information on the graph.Existing methods generally aggregate first-order neighbor information or node features on meta-paths based on message passing.However,less attention is paid to edge features and the overall semantic information of meta-paths,so the information of heterogeneous graph is not fully exploited.This thesis investigates edge features and the overall semantic features of meta-paths,proposes two models,and conducts a large number of experiments on widely used real-world datasets to verify the effectiveness of the models proposed in this thesis.The main works of this thesis are as follows:1.In the residual attention network method based on the aggregation of first-order neighbor information,this thesis proposes the attention node feature representation method using edge features RHGN.This method adds edge features to the aggregation of node feature information,adds an additional auxiliary task to determine whether the constructed triplet exists for updating the edge feature information.The proposed model performs better on multiple datasets,which verifies the effectiveness of the proposed method RHGN.2.In the attention network based meta-path method,this thesis proposes the node feature representation method based on meta-path HMAN.This method constructs a new local subgraph for the target node by sampling the meta-path instances according to the number of meta-path types,redesigns the inner meta-path feature aggregation layer and the inter meta-path feature aggregation layer to obtain the meta-path type features and the semantic features of the target node respectively.Finally,the node semantic features will be converted to the target node features.HMAN does not rely on prior domain knowledge and only requires adjusting the meta-path length.The proposed model demonstrates strong performance across multiple datasets,confirming the effectiveness of the proposed method HMAN.