Research On Community Discovery Algorithm Based On Network Representation Learning

With the expansion of network scale and the improvement of information level,more and more network analysis technologies,such as community discovery and link prediction,are used to obtain effective information from real social networks to describe the complex relationship between things.These valid information include not only the topology information of the network,but also other node related information.Node as an important research perspective,only by making full use of the characteristics of user nodes,the community structure detected by community discovery algorithm can accurately reflect the organization mechanism of the network.In order to extract the features of user nodes efficiently,it is necessary to fully express the relationship of nodes,which is also the reason for using representation learning technology to improve the effect of community discovery.In order to discover the community structure in social networks more accurately,this thesis studies the community discovery algorithm based on network representation learning.The following research has been done:1.To solve the problem of network sparsity,we need to use a node representation structure in line with the characteristics of social networks to improve the accuracy of feature representation;As well as the status of attribute information in social network research is becoming more and more important.In this section,we propose a network representation learning algorithm with attributes based on autoencoder.Firstly,the network with attributes is generated according to the information of core nodes and attribute nodes,and the relationship between nodes is described by meta structure;Then the node similarity is calculated and the final similarity matrix is constructed;Thirdly,the variable autoencoder is used to reduce the dimension of data,and the node representation vector is generated.So that each representation vector contains the structure information of the original node to the greatest extent.In the contrast experiment,the precision,recall,F1 Score,AUC and NMI value are taken as the evaluation criteria to verify the effectiveness of the proposed algorithm and the feasibility of applying it to the community discovery task.2.To solve the problem that community discovery needs to ensure the accuracy of feature representation of existing and new network nodes,as well as the limitation of modularity judgment and the need to improve the process of community partition,this chapter proposes a community discovery algorithm based on network representation learning and improved random walk.Based on the previous network representation learning,the node similarity is used to assign the network edge weight,and the probability transfer matrix is constructed;According to the joining time and meta structure style of network nodes,the method of selecting random walk target is improved;Then,the optimal representation vectors of nodes are generated by network representation learning,and their cosine similarity is calculated;In the stage of community discovery,cosine similarity is used to guide the generation of primary community,and local modularity and modularity change are used to judge the progress of secondary community merging;Through simulation experiments,the stability and effectiveness of the proposed method in community discovery are verified by comparing the modular optimization algorithm and the improved random walk algorithm.