Research On Ensemble Clustering Algorithm Based On Three-way Decisions

There are two steps to implement the ensemble clustering algorithm:Firstly,generating basic clustering members by basic clustering algorithm.Secondly,integrating basic clustering members by consensus strategy.Because ensemble clustering has an effective performance than classic clustering algorithms in data processing,ensemble clustering has gradually become a hot topic in the field of unsupervised learning.The current research on ensemble clustering mainly focuses on integration strategies,but the attention regarding the measurement and optimization of basic cluster is less emphasized.In this study,on the basis of information entropy theory,a quality metric for the basic clustering member is proposed.Moreover,based on the quality metric,a novel ensemble clustering algorithm framework and three basic clustering filtering algorithms are proposed.Specific research efforts include.(1)On basis of information entropy theory,using information entropy to measure the uncertainty of various clusters in basic clustering.Combined with the mutual information theory,the average cluster uncertainty of the basic clustering members relative to the basic clustering members set(member set)is obtained and defined as the quality of the basic clustering members,which is recorded as the cluster average entropy.(2)Extend the algorithmic framework for ensemble clustering by introducing a basic clustering pre-processing step.Combined with two branch decisions(a special expression of three-way decisions),three-way decisions and sequential three-way decisions,the framework of basic clustering pre-processing(further screening basic clustering)has been built.(3)According to the above three-step framework,two branch basic clustering filtering mechanisms(2BIA)based on two branch decisions and two basic clustering filtering mechanisms(BCF3WD and BCFS3WD)based on three-way decisions are constructed.Specifically,in 22BIA,if the quality of the basic clustering member is less than the preset threshold ?,the member is deleted and a new member is added to keep the cardinality of the member set unchanged.In BCF3WD,if the quality of the basic clustering member is less than the preset threshold ?,the member is deleted;if the quality of the member is greater than the preset threshold ?,the member is retained;if the quality measurement of the member is greater than a and less than ?,the member quality will be recalculated.Similar to BCF3WD,in BCFS3WD,three-way decisions are made first.Secondly,change the threshold ?and ? to make three-way decisions again.All above three mechanisms are repeated until the stop condition is reached.(4)Comparative experiments show that all three filtering algorithms are effective in improving ensemble clustering performance.For complex data sets,the three algorithms filter significantly,and the sequential three-way filtering method has less time-consuming than the remaining two algorithms.