| Heterogeneous social network is a social network containing different types of nodes and edges.It is different from the traditional homogeneous social network in that it contains a lot of semantic information.The link prediction on the heterogeneous social network refers to predicting the possibility of forming an edge between two nodes based on the topology and semantic information provided by the heterogeneous social network.The commonly used tool for extracting topology and semantic information in heterogeneous social networks is meta path,which is a sequence with nodes as end-points,nodes and edges appearing alternately.Therefore,the research focus of this paper is to use meta paths to extract and use topology and semantic information in heterogeneous social networks for link prediction.In recent years,researches have proposed many methods for extracting topolog-ical and semantic information in heterogeneous social networks based on meta paths for link prediction.They have achieved good performance in link prediction tasks.However,these methods have some shortcomings,and these shortcomings lead to their performance needs to be further improved.For example,they only use a single meta path,which leads to insufficient utilization of semantic information in heterogeneous social networks.At the same time,these methods focus on using the meta path to ex-tract features at the node level,but not using the meta path to extract features at the edge level,which makes their extracted features not well suited for link prediction,which is closely related to edges.In addition,these methods only consider the use of internal information of a single social network,which makes them unable to achieve good per-formance in sparse networks.In order to solve the shortcomings in the existing work and improve the performance of link prediction in heterogeneous social networks,the main work of this paper are as follow:Aiming at the problem of insufficient utilization of topology and semantic infor-mation of heterogeneous social networks by using only a single meta path in existing methods,this paper proposes a solution based on semantic subgraphs and graph at-tention mechanisms.This method can construct different semantic subgraphs based on different meta paths to extract different topological and semantic information in heterogeneous social networks.At the same time,the method also uses the attention mechanism within the subgraph to learn the weights of different nodes in the semantic subgraph,and uses the attention mechanism between the semantic subgraphs to learn the weights of different semantic subgraphs.Aiming at the problem that the existing research only uses the meta path to extract features at the node level,which makes these features not well suited for link prediction tasks,the solution in this paper uses a concept of semantic information flux based on meta path,which can use meta paths to extract features from the edge level.The method obtains the features closely related to edges by calculating the semantic information flux,and calculates the features related to different semantics based on different meta paths.Aiming at the problem that many existing studies only use the topology and se-mantic information within a single network,which cannot solve the problem of net-work sparsity,this paper proposes a solution based on cross-network meta paths.This method uses cross-network meta paths based on anchor links to transfer information from the source network to the target network to make up for the problem of insuffi-cient information in the target network.It uses these additional information to improve the performance of link prediction.And the information gain ratio is used to select the features extracted based on the meta paths to avoid the use of misleading information,which would affect the performance of link prediction.At the same time,a method based on AUC(Area Under ROC curve)optimization is used to solve the PU(Positive and Unlabled instance)learning problem of link prediction. |
