Research On Link Prediction Of Heterogeneous Networks Based On Graph Representation And Deep Learning

In the real world of the rapid development of information network,the network can well describe the complex relationship between different entities,in which each entity can correspond to the node in the network,and the edge in the network represents the relationship between the two entities.Complex network is the largest network in real life.Its structure is more complex and has rich data.Mining and analyzing the information to obtain more valuable information has attracted extensive attention of scholars.Link prediction is one of the challenging tasks in the research of these networks.The main task of link prediction is to predict whether a new relationship may occur in the network,that is,to predict the possibility of two nodes without edge connection in the network.The representation of networks in the real world can generally be divided into two categories:homogeneous networks and heterogeneous networks.There are only one type of nodes and edges in homogeneous networks,while there are many types of nodes and edges in heterogeneous networks.Most of the research at this stage choose homogeneous networks,but most of the networks in real life are heterogeneous networks,so the use of heterogeneous networks is closer to life and has more practical research significance.Heterogeneous information networks are rich in node and edge types,which makes the research of link prediction under heterogeneous networks very difficult.By combining the existing heterogeneous network research and network embedding methods,this paper proposes a heterogeneous network link prediction algorithm based on graph representation learning.The main contents are as follows: a method of dividing heterogeneous networks based on multichannel networks is proposed;A deep relational feature learning model graph conception model with improved conception model is proposed for graph convolution;The above method of dividing heterogeneous networks and graph convolution graph inception model are combined into mdgcn(multi channel deep graph convolutional network)model,which is applied to link prediction.1.Method of dividing heterogeneous networks based on multi-channel networks: the types of nodes and edges in heterogeneous networks are diverse,and the characteristics of different types of nodes and edges are different,which makes the relevant models of deep learning unable to be directly applied,so the convolution operation becomes very difficult.For example,in traditional GCN(graph convolutional network),all nodes share the same convolution kernel,which can not be applied in heterogeneous networks.Therefore,this paper proposes a multichannel network.Each channel of the network is a homogeneous network,and the nodes in each channel are connected through a specific type of meta path.Among them,each meta path represents the complex relationship of specific types of nodes,which can be defined as a specific edge type of the network.By learning the node embedding of each channel,and finally splicing,this paper realizes the information acquisition of heterogeneous networks.2.Improved the deep relational feature learning model graph perception model of perception model: GCN(graph convolution network)has developed very rapidly in recent years,but there are many problems in the research of depth GCN at this stage.The increase of the number of layers of graph convolution will cause the gradient to disappear,that is,the back propagation of the network causes the features to be too smooth,which makes the final features of nodes converge gradually and finally become the same value.This means that the number of layers of GCN generally does not exceed 4.The concept model proposed in CNN can only deal with Euclidean data,but can not deal with graph data.Therefore,this paper uses graph concept model to deeply learn the relationship features of graph data.The model uses the method of multiple convolution kernels to extract the relationship characteristics of sub networks by combining multiple convolution kernels of different sizes,so as to increase the depth of the network.3.Mdgcn model combining multi-channel network division heterogeneous network and deep relational feature learning model graph perception model: traditional GCN can only carry out some shallow networks,and can not deal with heterogeneous networks with diverse nodes and edges.Therefore,this paper processes the heterogeneous network information through the multi-channel network division,then obtains the deep relationship characteristics through the graph inception model,finally becomes the edge vector through the Hadamard distance processing,and calculates the probability of edge existence through logical regression to complete the link prediction task.The experimental data sets selected in this paper are three real heterogeneous network data sets: Cora,DBLP and IMDB.The selected evaluation indexes are AUC and precision.Through comparison with traditional algorithms PA,LP,Katz,LRW,SRW and traditional GCN,the results show that this method has a certain effect in dealing with the diversity of heterogeneous network nodes and edge types,and can successfully obtain the deeper relationship characteristics of nodes.Compared with traditional algorithms,it has a certain improvement in link prediction accuracy,In the three data sets,the highest AUC value increased by 13.02%,and the lowest AUC value also increased by 1.96%.