Study Of Mutigranular Attribute Reduction Method Based On Clustering

At present,with the rapid development of science and technology,a large amount of unlabeled data has been produced in practical applications.The clustering method is a representative method for processing such unlabeled data.However,due to the existence of redundant information,traditional clustering methods suffer from large time consumption and low accuracy.On the other hand,attribute reduction based on rough set theory can reduce redundant attributes and extract useful information while maintaining the same identification ability as the original information system.This paper proposes a cluster-based mutigranular attribute reduction method for unlabeled data.Through the adjustment of the K value in the clustering algorithm,the mutigranular calculation is performed to form the partitions of the universe from coarse to fine.The clustering results are used to complete the supervised attribute reduction in the rough set theory to remove the redundant attributes,and finally the KNN algorithm is used to determine classification results.Specifically,the main work of this thesis is divided into two parts:On the one hand,for equivalence relation-based information systems composed of symbolic data,this paper uses K-modes clustering algorithm,and then uses the clustering results as class labels.By adjusting the K value to form multiple divisions of the universe of domains,the positive domain dicernibility matrix is used to reduce the redundancy attribute for each division.Then the dimensionality of datasets is reduced and the algorithm cost can be saved.On the other hand,for information systems with ordinal attributes,we use dominance relation-based rough set method to reduce redundant attributes based on the mutigranular computation of K-means clustering.Finally,the proposed method is compared with the traditional rough set model and the traditional clustering method.Because the proposed method performs supervised attribute reduction based on clustering information,it improves the unsupervised attribute reduction method.Additionally,since redundant information has been reduced,the performance of clustering algorithm is also improved.