Adaptive K-means Algorithm Based on Three-Way Decision

Abstract

AbstractThe focus of traditional k-means and its related improved algorithms are to find the initial cluster centers and the appropriate number of clusters, and allocate the samples to the clusters with clear boundaries. These algorithms cannot solve the problems of clusters with imprecise boundaries and inaccurate decisions due to inaccurate information or insufficient data. Three-way clustering can solve this problem to a certain extent. However, most of the existing three-way clustering algorithms divide all clusters into three regions with the same threshold, or divide three regions subjectively. These algorithms are not suitable for clusters with different sizes and densities. To solve the above problems, an adaptive k-means algorithm based on three-way decision is proposed in this paper. First, the traditional clustering results are taken as target set and core region. The distance between each sample in the target set is used as the candidate neighborhood radius threshold. At the same time, neighborhood relationship is introduced to calculate the accuracy of approximation, upper and lower approximation of the target set under the current neighborhood relationship. Second, a boundary control coefficient is defined according to the accuracy of approximation, and as many abnormal data as possible are classified into boundary regions to transform traditional clustering into three-way clustering adapted to different sizes and densities. Finally, five indexes are compared on UCI data set and artificial data set, and the experimental results indicate the effectiveness of the proposed algorithm.KeywordsThree-way clusteringThree-way decisionNeighborhoodK-meansAccuracy of approximation