Research On Fuzzy Rough Support Vector Clustering Method

With the rapid development of intelligent technology,the data scale of all walks of life is increasing in geometric series.More and more studies focus on how to explore potentially useful information from massive data.Cluster analysis is an important branch of data mining.Support vector clustering(SV)is a cluster analysis algorithm based on contour.The SVC principle is through building standards,setting the sample points that describe the cluster contour in the original data space.These sample points are called support vector.Then,SVC according to some strategies marks the support vector.The support vector with the same label is enclosed a cluster in the original data space.Although SVC has been widely applied,it still has lots of problems,such as the cluster overlap,parameter sensitivity,high time complexity and so on.On the basis of classical support vector selection strategy,by considering the local density of the data and introducing the fuzzy rough set theory model,two support vector selection strategies are proposed.The first one is accomplished by considering the local density.Another one incorporates the theory of fuzzy rough set.Based on the spectral clustering labeling method(SCLM),by taking into account the close proximity of data in the original data space,modified spectral clustering labeling method(MSCLM)is proposed.And on the basis of this,a new labeling method with low complexity(NLM)is proposed.For select support vector to calculate redundant problem,the local density of each sample point is calculated and an adaptive local density threshold is identified.Then selecting the samples that local density less than or equal the threshold value conducts support vector alternate set.Then support vector are selected from it.For data set with cluster overlap and inhomogeneous distribution of sample inside cluster,a fuzzy rough set model is proposed,by calculating approximation of fuzzy rough sets for each sample point.SVSSFRS make samples between clusters are judged to boundary support vectors,and in the cluster the sparse sample are selected as support vectors.Theoretical analysis shows that the support vectors determined by this method are closing to the same cluster and far from different clusters in the original data space.In the high dimensional feature space,support vectors of the same cluster have small angle and support vectors of the different cluster have big angle.For the result of FCM clustering cannot reflect the distribution of the original data space,a modified spectral clustering labeling method(MSCLM)is proposed by introducing the neighbor relationship in the original data space,following up the result of clustering,improving the accuracy of support vector markers and the precision of the clustering.In order to reduce the time complexity,a simplified support vector marking method on the basis of the spatial relationship between support vectors is proposed.The proposed MSCLM and NLM make the experiments.Experimental results show effectiveness of two methods.The proposed two method is compared against SCLM and other commonly used labeling method.Experimental results show superior performance of two methods.