Research On Community Detection Methods In Complex Netwokrs

With the continuous development of society and technology,the social relationships and contact methods between individuals become more and more complex,and a lot of complex systems have been formed.It is difficult to obtain valuable information directly from complex systems.In order to solve this problem,complex systems are abstracted as complex networks which are relatively simple in form.The general rules and characteristics of complex systems can be effectively excavated by researching deeply on complex networks.At present,some characteristics of complex networks have been found such as small-world,scale-free,community structure and so on.The community is composed of nodes which are closely related.The most important feature of community structure is that links between nodes in the same community are relatively tight,and links between nodes in different communities are relatively sparse.People can easily access to information that they are interested in with community structure which could be applied to many other fields such as hot spot mining,personalized recommendation,link prediction and so on.As a result,community detection methods have become a focus of complex networks.In this paper,we study on the specific problems of community detection methods in complex networks and give the corresponding solutions in four aspects: improving stability of community detection results,overcoming the lack of network information with known background information,mining overlapping community structure,combining content topics of nodes and relationships of links.First,a community detection method based on stable label propagation is proposed to improve the stability of community detection results aiming at the instability of traditional label propagation community detection methods.The randomness of traditional community detection methods based on label propagation causes a lot of differences between the results in the same network and reduces the quality of the results.Aiming at dealing with this,this paper proposes a community detection method based on stable label propagation.Firstly,labels are initialized by searching for non-overlapping triangles in the networks.Then nodes queues are formed with labels entropy for reducing randomness.At last,new labels are chose by the distribution of adjacent nodes' labels when maximum labels are not unique.Second,a semi-supervised local clustering community detection method is proposed to overcome the problem of missing network information by using some known background information aiming at deviations of community detection results caused by partial networks information loss.In traditional community detectionmethods,the known background information is not taken advantage of to overcome the loss of partial networks information which causes deviations of community detection results.Aiming at dealing with this,this paper proposes a semi-supervised local clustering community detection method.First,it improves traditional measurement of how nodes to join the community,making the community detection results more in line with the definition of community structure.Then,the reward and punishment measures are used to encourage formation of the community structures which are consistent with known information and to prevent formation of the community structures which are not consistent with known information.This method can dig out the original community structure better in the networks with incomplete information.Third,overlapping community detection methods based on semi-synchronous label propagation and local clustering are proposed to improve the quality of overlapping community structure detection aiming at the problem of nodes' role diversity and community structure hard division.Each node belongs to only one community in traditional community detection methods,which ignore the overlapping community structures that are more in line with the actual situation.Aiming at dealing with this,this paper proposes a semi-synchronous label propagation overlapping community detection approach.This approach combines asynchronous propagation strategy and synchronous propagation strategy.While avoiding the oscillation problem of label propagation,it achieves desirable balance between the computational efficiency and the quality of community detection results.In addition,an overlapping community detection method based on local cluster is proposed in this paper.The improved Page Rank algorithm is used in seed node selection,and Spin-glass model is used as criteria for node selection during community expansion.Overlapping community structure in complex networks can be better discovered by these methods.Finally,a community detection method of combining content topics and relationships of links is proposed to get the community detection results that are more accord with the real situation aiming at ignoring the attribute node level content information in traditional community detection methods.In traditional community detection methods,only links between nodes are considered to discover community,ignoring content attributes of nodes.Aiming at dealing with this,this paper puts forward a community detection method which combines content topics and relationships of links.First of all,LDA model is used to access to content topics vectors of nodes,and their similarities.Then,the similarities are fused into label propagation and local clustering for community detection.In the method of combining content topics and label propagation,candidate labels are endowed with weight values which are the similarities of content topics vectors between candidatenodes and propagated node.The new label of propagated node is the candidate label with maximum weight value.In the method of combing content topics and local clustering,similarities in content topics and relationships of links between candidate nodes and current community are integrated,and the node with the largest increasement in community density is selected to join current community.Comparing with community detection methods only based on content topics or relationships of links,this method has achieved more in line with the actual situation of community detection results.