Research On Multi-layer Network Community Collaborative Detection

With the rapid development of data acquisition technology,there are more and more large-scale,diverse and complex data.If we can analyze and process the data effectively and find the potential internal relationship between the data,it will be of great significance to the research of related fields.At present,most of the data in the real world can be expressed as different individuals and the relationship between individuals,so it can be mapped into the form of network.The nodes in the network represent individuals,and the edges represent the connections between individuals.Community detection is an important research direction in network analysis,which aims to divide data points in complex networks into different clusters,so that the connections within clusters are as dense as possible,and the connections between clusters are as sparse as possible,so as to dig out the potential community structure in complex networks,make people better understand complex networks,and study the network structure,function and characteristics,so it is widely used in the related research of large data and artificial intelligence.At present,many scholars have studied the problem of community detection.These works mainly focus on the single-layer network,that is,there is only one connection form between the nodes of the network,but the network in the real environment may have multiple connection types at the same time.For example,in social networks,individuals will contact each other by telephone,SMS,microblog and other means.Only by obtaining the information of all platforms can we accurately judge whether two individuals in the network belong to the same community.In order to cover the multi type characteristics of these relationships,through the same node set and hierarchical description of different connections to form a multi-layer network,and community detection is obviously of stronger scientific significance,so it has attracted extensive attention in network science in recent years.However,the existing methods only fuse the connections of multi-layer network into single-layer network,without considering the importance of different connection types for community detection,which seriously interferes with the community structure of the fused single-layer network.In addition,there are great differences in connections between layers in multi-layer networks,and the multi-layer structure may further aggravate the sparsity of connections within layers.At present,there is no robust community detection strategy in multi-layer networks to systematically mine the network structure.In order to solve the above problems,this paper makes an in-depth study on community detection in multi-layer networks,and achieves the following research results:(1)a new collaborative representation algorithm in layers is proposed to identify communities in multi-layer networks.Considering that the noise difference of different connections in multi-layer network will lead to the uneven performance of community detection,this algorithm proposes an adaptive layer weighting strategy,which constantly updates the weights of different layers to describe the importance of each layer network to community detection.In addition,the algorithm is based on non-negative matrix factorization,and forces the low dimensional representation vectors of adjacent layer networks to be close to each other to obtain consistent low dimensional representation,so as to reveal the community structure shared across layers.In addition,we also design a related optimization method to optimize the decomposition factor and the weight of different layers at the same time by alternating iterations until convergence.A large number of experiments on various types of complex networks including citation,social,economic and biological networks show that this method has good community detection accuracy.(2)A subspace based community detection algorithm for multi-layer networks is proposed.In view of the sparsity and connectivity difference of multi-layer networks,this method extends the idea of sparse subspace to multi-layer networks to improve the robustness of community detection,and integrates the low dimensional representation similarity constraints between different layers as distance regularization terms and non-negative constraints into the sparse subspace clustering framework,so that it can make use of the global and local information of the data for graphics.Furthermore,a novel sparse constraint is introduced to promote the learning graph to have a clearer clustering structure.To solve this problem,we specially design an effective iterative algorithm to optimize the framework.The experimental results on multi domain datasets show that this method can significantly improve the community clustering performance of complex networks.