Research On Clustering Algorithm For Complex Network

Numerous complex systems in real world are abstracted into complex networks for research and analysis,which has become an effective research method in network science.Cluster structure has been proved to be a ubiquitous structural feature on complex networks.Mining the cluster structure of the network not only plays an important role in revealing the various information,functions and changing characteristics hidden in complex systems.At the same time,it also has a certain guiding significance for understanding various interactive behaviors in the real world,such as personalized recommendations for social relationships,predicting the interactive relationships between different proteins,and mining social media communication behaviors.Network clustering algorithm is one of the effective methods to identify complex network cluster structure.How to design efficient clustering algorithm to mine network cluster structure has become one of the hotspots of current network science research.In the real world,the relationship between the entities of some complex systems is relatively simple,which can be studied and analyzed by constructing a single layer network.With the development of the times,the relationships between entities in most complex systems have gradually diversified,and are no longer just connected through a single relationship.It has become an effective method to research and analyze these complex systems by constructing a multi layer network.Therefore,constructing single layer networks or multi layer networks for different complex systems,and then developing clustering algorithms to mine their cluster structure has become the main research content of the paper.(1)Aiming at the problem of poor clustering quality caused by unbalanced exploration and exploitation of multi objective clustering method in single layer network clustering,the paper designs a multi objective discrete moth flame optimization algorithm with paradigm constraints(DMFO).The proposed method effectively balances the exploration and development process when searching for the population,thereby obtaining a better network cluster structure.First,the single layer network clustering problem is abstracted into a multi objective optimization problem.The classic Kernel K means(KKM)and Ratio Cut(RC)are selected as a pair of contradictory objective functions for optimization,and modularity is used as an evaluation index for selecting offspring populations.Secondly,in order to balance exploration and exploitation of the population in the solution space search,DMFO redesigned the SFG and SFS processes to ensure the local search ability and the global search ability,respectively.Finally,a Tchebycheff decomposition approach with an l₂ norm constraint on the direction vector(2 Tch)is used to decompose the multi objective optimization problem into a set of scalar optimization sub problems for solution,which can obtain a better Pareto set.Numerous simulation experiments on benchmark networks and real network datasets show that the proposed DMFO algorithm has superior performance compared to other algorithms in solving single layer network clustering.(2)Aiming at the problem that traditional network clustering methods cannot accurately cluster on extremely sparse or complex multi layer networks,a semi supervised method based on consensus subspace graph regularization(S MCGR)is proposed.The proposed method realizes the sufficient supervision of the consensus prior information on the optimization of the common low dimensional representation matrix,thereby effectively solving the problem of clustering difficulties in the current clustering algorithm on the multi layer network with a lot of noise.First,the consensus prior information of the network is obtained through the non overlapping greedy search method,and the consensus prior matrix is constructed.Then,the graph regularization term encoding the consensus prior information is constructed to supervise the consensus subspace fusion of the multi layer network.Finally,the consensus prior matrix is used to preprocess and modify the adjacency matrix of each network layer,so that a better low dimensional representation of each layer can be obtained for consensus subspace fusion.Based on the above strategies,S MCGR method can make full use of the obtained consensus prior information to supervise the network clustering,so as to improve the network clustering performance.Extensive experiments on synthetic and real world multi layer datasets show that the proposed S MCGR method has better accuracy,applicability,robustness and lower computation complexity.On the whole,this paper focuses on solving single layer network and multi layer network clustering problems by developing a multi objective natural heuristic algorithm based on constraints on direction vectors and a semi supervised method based on consensus subspace graph regularization,respectively.Extensive simulation experiments verify that the two proposed algorithms have outstanding network clustering performance.