Research And Application Of Gated Graph Attention Network

In real world,many scenes are represented by more general graph-structured data,such as citation network,protein molecular network,knowledge graph and so on.The traditional convolutional neural network is powerless in mining the semantics and patterns of graph-structured data.In order to perform convolution on graph-structured data,a graph convolutional neural network is proposed to solve the problem caused by the lack of translation invariance on graph-structured data.In recent years,graph convolutional neural network has developed rapidly and has been successfully applied in many fields,such as user recommendation system,public opinion monitoring and control,and cancer prediction.After studying the network structure and development process of graph convolutional neural network in detail,this dissertation summarizes two main challenges of it:lack of multi-hop neighborhood information;and lack of prior knowledge enhancement.On this basis,a graph attention convolutional neural network based on gating mechanism is proposed and applied to graph node classification.Then this model is modified adaptively according to the characteristics of cancer classification and link prediction,and two new models are proposed.Finally,the validity and interpretability of the proposed models are verified.The three models proposed in this dissertation are as follows:1.Gated Graph Attention Network(GGAT)is proposed and applied to graph node classification.In GGAT,a gating mechanism is introduced to break through the limitations of the mainstream graph neural networks that can only utilize short-range information,which can also filter information selectively.Experimental results on benchmark graph datasets Cora,Citeseer and Pubmed show that GGAT achieves higher accuracy than traditional neural network models in graph node classification.2.Prior knowledge is aggregated into GGAT(GGAT with Prior Knowledge,PKGGAT)and it is applied to cancer classification.Experimental results show that original GGAT achieves higher accuracy than traditional machine learning algorithms and neural network models in cancer classification;After the fusion of prior knowledge,PK-GGAT achivevs higher accuracy than GGAT,which shows that prior knowledge can enhance the performance of the model.In addition,a hybrid feature selection algorithm is proposed to reduce the training time of traditional machine learning methods used in this dissertation.3.Relational Gated Graph Attention Network(R-GGAT)is proposed and applied to link prediction.The experimental results on knowledge graphs FB15k,FB15k-237,UMLS and WN18RR show that R-GGAT achieves higher accuracy than the traditional neural network models.