Research On Overlapping Community Detection Based On Network Structure Embedding

In real life,complex systems in varieties of fields,such as social media,biology,transportation logistics and so on,can be represented as the form of complex networks.Community structure refers to the character that nodes in one set are closely connected and the connection between nodes in different sets are relatively sparse.The concept of community in network is extremely broad.For instance,it may represent a club with common hobbies,or staffs of a certain department in the company.In many cases,the nodes in complex systems often simultaneously exist in multiple communities,that is,overlapping community.As a momentous characteristic of network structure,overlapping community detection has always been a significant direction in complex network research field.Recently,numerous algorithms about overlapping community detection in complex network have been proposed.However,some deficiencies in these algorithms still exist.For example,many overlapping community detection algorithms are unsatisfactory in network with more complicated network structure;the overlapping community detection algorithms based on label propagation have unstable community partition results.In response to above questions,the primary work of this thesis is as follows:1.The algorithm,overlapping community detection based on network embedding and node density(OCDNED),is proposed.The loss function is constructed based on the first-order proximity and second-order proximity in network.To avoid the deficiency of the traditional node representation,such as high spatial complexity or difficulty in capturing the similarity between distant nodes,the node representations with low-dimension are obtained through network structure embedding and the local structure characteristics of network is preserved.Meanwhile,the algorithm improves the node's allocation strategy of community in density peaks by combining with belonging coefficient vector of community labels and controls the propagation degree of labels through parameters ? and ?,so that communities with different overlapping degree can be found.Comparative experiments on multiple real-world network datasets and LFR benchmark datasets show that the algorithm not only performs well in networks with significant community structure,but also properly find overlapping communities in network with more complicated community structure.2.The algorithm,overlapping community detection based on node centrality and threeway decision(OCDNC?3WD),is proposed.Most community detection algorithms based on heuristics adopt the same allocation strategy for all nodes in the network during community partition.Considering that nodes in network are not uniformly distributed,different nodes are distributed in different positions.Based on the idea of three-way decision,the nodes in network are categorized into three types: community core nodes,undetermined nodes and isolated nodes.Different types of nodes are served by different strategies.In the process of node label distribution,the centrality of nodes and the similarity between nodes are simultaneously considered,which not only improves the efficiency of community partition,but also ensures the stability of community discovery.Experiments on multiple network datasets show that the OCDNC?3WD algorithm can accurately detects overlapping communities in network structure.