Research On The Key Technology Of Heterogeneous Network Alignment

With the rapid development of science and technology,the Internet has gradually become an indispensable tool in people's daily lives.It is a powerful graph structure that can naturally capture the relationship between objects in massive data.In order to obtain more hidden information about the object,many graph mining tasks need to combine multiple network data for analysis.However,the primary problem faced by data aggregation in heterogeneous networks is "how to align objects of different platforms".Many downstream tasks are constructed after objects are aligned.For example,pattern matching of protein networks,user identification in social networks,cross-domain personalized recommendation and social link prediction,etc.Therefore,network alignment is of great significance for a wide range of applications.This paper has done the following two tasks for the alignment of heterogeneous networks:(1)At present,the solutions to most heterogeneous network alignment problems can be summarized into the following three steps: feature extraction,model construction and node matching.The first step is to extract the available information from the node's attribute information and network structure as the features of the node,which identifies the identity of the node.In the second step,statistical or vectorized representation of the features extracted in the previous step is used as input for the model construction stage.Then,a supervised or unsupervised model is trained according to the availability of anchor node pairs(ie,pre-known alignment node pairs).Finally,the trained model is used to predict whether the identities of the two nodes match.Currently,most network alignment methods only focus on the local structure or global structure of nodes in the feature extraction stage.This feature extraction method causes the lack of node information to a certain extent.Therefore,this paper proposes a multi-scale modeling mechanism,which not only builds the features of nodes on the global structure,but also strengthens the expressive ability of node information from the local structure,and enhances the richness of node features.At the same time,in the model training stage,we often face problems of high computational complexity.In order to reduce the computational complexity of the model training process,this paper introduces an implicit matrix decomposition method based on QR decomposition,which uses a small number of nodes to achieve the ability to express all nodes,which greatly reduces the computational complexity.(2)The implicit matrix factorization method often needs to filter out a small number of nodes,and use these nodes to establish an association relationship with all nodes.These selected small numbers of nodes are called landmark nodes.At present,models that use implicit matrix factorization to solve network alignment problems often use random sampling technology to select landmark nodes.This screening method treats all nodes equally,ignoring the particularity and importance of landmark nodes.Therefore,this paper proposes a sampling strategy based on node importance.According to the importance of nodes,this strategy selects the most representative nodes from all nodes as landmark nodes,and then builds their association with global nodes.The heterogeneous network alignment algorithm proposed in this paper can be widely used in different fields,such as social network user matching,link prediction,user portrait,friend recommendation,protein analysis,pattern recognition and so on in practical problems.In order to verify the effectiveness of the two algorithms proposed in this article,a large number of experiments were done on three different types,different scales,and different sparsity data sets,and three different evaluation indicators were used to compare the model from the perspective of prediction and ranking to evaluate.This kind of evaluation index evaluates the model from the perspective of prediction and ranking.Experimental results show that the two network alignment algorithms proposed in this paper have better performance.