Research On Analysis Of Heterogeneous Network Structure Based On Robust Graph Representation

With the development of society,science and technology,people have accumulated a lot of data.These data types are diverse and have different structures,so it is urgent to make a model and analyze them effectively.Heterogeneous network is a kind of information network which contains many kinds of nodes and many kinds of links between different types of nodes.It can organize all kinds of data organically and form a graphical data model,which has significant influence on research.Drug information network is an important heterogeneous network,which can integrate various types of biological/drug factors and simulate the complex relationship among them.Therefore,the analysis of biological/drug information has become a hot topic.In current drug information analysis,drug-targeting interaction identification is an important step in drug discovery and adverse reaction prediction.In the past,we use biochemical experiments to detect the relationship between drugs and targets,so as to determine whether a specific drug can play a role in a certain target,but this traditional biochemical method will cost a lot of manpower and material resources.In the real biological environment,the relationship between drug and target and themselves constitute a complex heterogeneous network,which profoundly affects the prediction performance of drug-target interaction.With the rapid development of the computer field in recent years,and the continuous design and demonstration of models to explore the relationship between drugs and targets,we have been able to use emerging computer technologies to help people more effectively explore a wider range of drug action areas.However,existing approaches typically rely only on drug-target interactions,which can be very sparse and have a bad influence on drug-target prediction.To solve this problem,this paper will build two different models to analyze the information between drugs and targets.First,to adapt to the sparse connection characteristics of drug-targeting networks,we propose a dual graph regularization robust principal component analysis in heterogeneous networks.The model aims to decompose the bipartite graph of drug-target interaction into two low-rank matrices,which represent the potential representation of drugs and targets,respectively,and also to smooth drug-drug similarity network and target-target similarity network.In addition,an improved robust PCA model is used in the decomposition phase to suppress the extensive noise linkage,thus approximating the existing sparse drug-target network more accurately.In order to optimize the model,we also designed an effective alternate iteration algorithm,in which the subproblems at each stage have closed solutions,which significantly improves the computational efficiency.Finally,a large number of experiments were carried out on a real drug-target heterogeneous network,verifying that the proposed method has more robust prediction results.On the other hand,with the continuous progress of biotechnology,people have known many types of drug relationship data and target relationship data.Using this multi-view data is helpful to make up for the lack of traditional drug-target data and alleviate the problem of network sparsity.Based on the above analysis,this paper also designs a drug-target low-rank embedding method based on multi-view fusion.Based on multi-source heterogeneous data,the method retains the integrity of the original data to the maximum extent,reduces the cost of operation,and realizes the analysis and research of the relationship between drugs and targets.The model can be roughly decomposed into two steps.In the first step,multiple drug-drug relationship data and target-target relationship data are respectively fused to obtain the characterization of drug and target respectively.The second step is to find low-dimensional space to embed the relationship between drugs and target based on the results of the previous step,find out the feature vectors that conform to the model,and combine the expression of data in the low-dimensional space.Finally,this low-rank expression is used to predict the existence of drug-target relationships.We have carried out a large number of test experiments in drug-targeting data to verify the effectiveness of this method.