Research On Heterogeneous Graph Embedding Method And Application Based On Graph Neural Network

Heterogeneous graph embedding aims to project the inherited heterogeneity of the target graph along with the structural information into a vector in a low-dimensional space,so as to improve downstream tasks such as node classification,node clustering,recommendation systems,etc.As the graph neural network has achieved excellent results in the field of homogenous graph embedding,researchers began to combine heterogeneous graph embedding method with graph neural network,by utilizing the powerful representation ability of graph neural network,these model can capture the information in the graph effectively,which can improve the accuracy of downstream task prediction.Heterogeneous graph embedding methods usually rely on metapaths to model the unique heterogeneous structure information and semantic information in the graph.However,the existing metapath-based heterogeneous graph embedding methods usually assume that the given metapaths are mutually independent,which overemphasizes the local structure of each metapath,while ignoring the global correlation between them,so the captured heterogeneous structure information is limited.On the other hand,thanks to the unique structure of heterogeneous graphs and the excellent performance of heterogeneous graph embedding methods on graph analysis tasks,some researchers try to use heterogeneous graph embedding in natural language processing and recommendation systems.However,click-through rate prediction,as one of the most important tasks in the recommendation field,has rarely seen the application of heterogeneous embedding methods.Address to these limits,this paper uses heterogeneous graph embedding as the research foundation,combined with the rapid development of graph neural networks in recent years,and proposes the following two innovations:Firstly,existing heterogeneous graph embedding models usually assume that the given metapaths are mutually independent,therefore these models first generate corresponding node embeddings for each metapath,and then merge them.However,they overemphasize the local structure of each metapath,while ignoring the global correlation between them.Address to this limit,a Metapath-based Relational Selection Graph Neural Network(MRSGNN)is proposed.MRSGNN first transforms the original graph into a metapath-based multi-relation one.In order to capture the global correlation mentioned above,each relational selection graph convolution layer aggregate neighbor information under each relation in the generated graph directly,and rescale them based on attention mechanism.On the node classification task of three real-world heterogeneous graph datasets,MRSGNN achieved 94%,92%,61% prediction accuracy respectively,which consistently outperforms the state-of-the-art baselines.Subsequently,this paper combines the heterogeneous graph embedding with the click-through rate prediction task.The existing click-through rate prediction model projects features under different fields to a unified feature space to construct high-level feature interactions,which ignores the semantic differences of different fields and the correlation between features under the same field.Therefore,this paper proposes to introduce heterogeneous graphs into the click-through rate prediction task,by transforming input features into heterogeneous feature graphs,and using the proposed Feature interaction via Heterogeneous Graph Attention Network(Fi-HGAN)to model feature interaction from different fields along with the semantic differences between the fields.On the clickthrough rate prediction of two public benchmarks,Fi-HGAN achieved 0.8084 and 0.7766 AUC score,respectively,which consistently outperforms the state-of-the-art baselines.