An improved K-means clustering algorithm for two half-moon classification

Abstract

Classification problems of machine learning use supervised learning under specific targets to classify new observations. This work presents a new clustering and classification approach that combines an evolutionary algorithm with the K-means algorithm. In order to assess the performance of the proposed approach, the authors conducted a simulation study using a well-known benchmark problem called “two half-moon rings classification”. The selected problem introduces further complexity and higher classification challenge when a new observation is located in region of intersection of the two half-moons. The Cartesian coordinates of several points are used as a data set for two half-moon rings. The set is injected with complex overlap situations to constitute data points that belong to more than one class (ring) at a time. The modified set is investigated using the proposed clustering and classification approach. The proposed algorithm obtains the optimal cluster centers using genetic algorithm. Furthermore, it adopts whitening method to overcome the effect of overlapped points on clustering accuracy. Obtained classification results showed enhancement over those produced by the conventional K-means clustering algorithm. The results are consistent under different ring dimensions, and several overlap situations.