Studies On Group Detection And Its Propagation Control On Online Social Networks

Online social networks are important channels and platforms for people to communicate and disseminate information.It has both great positive effects and many negative effects on people's lives.Therefore,the online social networks need to be studied and analyzed to enhance their positive effects and eliminate their negative effects.The online social network group composed of many individuals with some common characteristics is an important meso structure of online social network.Compared with communities that are defined based on the network connection density,the definition of network groups places more emphasis on the common characteristics of aggregated individual.The study of online social network groups is one of the basic work of online social network analysis.This research is aimed at the national Internet public opinion analysis and guidance,online group social interaction and other national security and social development demands.It deeply and systematically studies the social network group detection and its propagation control technology.Group detection refers to the process of detecting a set of groups based on certain features of the groups.At present,the structure features of the groups have been studied extensively,but the attribute features that reflect the aggregation factors of the groups are still lack of in-depth study.When detecting the groups,how to flexibly apply the various features of the groups,how to make full use of the network structure and attribute information and how to find the target group set which satisfies the specific application requirements are the problems that need to be solved urgently.Groups and information propagation are closely related,how to control the spread of negative information based on the selection of opposing source groups is also the current research hot and difficult problem.In view of the above problems,based on the analysis of structure and attribute features of group and its definition,this dissertation use a variety of optimization techniques to study the overlap and multi-level group detection methods based on multiple group features,the group detection method based on the integration of structure and attribute,target group detection method based on sample information,and further study the group-based rapid and effective network public opinion propagation control technology.The results of this dissertation will help to visualize the social network organizational structure,control the propagation influence of malicious behavior groups,and provide support for applications such as product recommendation and marketing,etc..The main innovations of this paper are as follows:(1)Aiming at the problem of overlap group detection and multi-level group detection under the circumstance that the network structure is provided while the individual attributes are not provided,this dissertation proposes an overlap group detection method based on the maximal clique and a multi-level group detection method based on multi-objective optimization,respectively.The overlap group detection method utilizes the maximal clique to help determine the group identity of the individuals.Experiments show this method can detect the overlap individuals and the overlap groups more accurately.The multi-level group detection method uses the multi-objective evolutionary algorithm that integrates local search technology to optimize two objective functions that describe the different properties of the groups simultaneously.This method is more accurate than the existing multi-objective optimization group detection methods.(2)Aiming at the problem of attribute homogeneous group detection when the network structure and the individual attributes are both provided,this dissertation proposes a group detection method combining structural cohesion and attribute homogeneity based on multi-objective optimization.This dissertation defines a homogeneity function to measure the homogeneity degree of the group and proves that the homogeneity function is maximized when the attribute values of different groups are different and those of the same group are the same.The multi-objective optimization evolutionary algorithm is used to optimize the modularity and homogeneity simultaneously.Experiments show the detected groups are cohesive and homogeneous and that different group structures correspond to different balances of the relative importance of structure and attribute.(3)Aiming at the problem of detecting the set of groups with specific features based on sample information for some applications,this dissertation proposes two target group detection methods based on sample information.When the sample information is two sample nodes,a sample information extension method is designed to extend two sample nodes to a set of exemplar nodes.The target subspace is inferred based on the exemplar nodes and then the target group set is mined.When the sample information is a set of sample attributes,a unified quality function is defined to measure the target subspace and group quality,and then each target subspace and its group is iteratively optimized.Compared with unsupervised group detection methods,the proposed methods can make full use of the sample information to quickly detect a set of target groups that meet the application requirements.(4)Aiming at the problem of how to quickly and effectively control the negative influence propagation initiated from any negative source group,this dissertation proposes a group propagation control method based on the local influence computation.This dissertation studies the group influence blocking maximization problem on two typical competitive propagation models.The idea of two proposed methods is to approximate the negative influence of any node on the local maximum influence tree of each node,to select the node with the maximum negative influence as the positive source,and update the blocked negative influence of related nodes.The proposed methods are several orders of magnitude faster than the greedy methods and have the similar influence blocking performance to greedy methods.