An improved K-means algorithm for big data

Abstract

An improved version of K-means clustering algorithm that can be applied to big data through lower processing loads with acceptable precision rates is presented here. In this method, the distances from one point to its two nearest centroids were used along with their variations in the last two iterations. Points with an equidistance threshold greater than the equidistance index were eliminated from the distance calculations and were stabilised in the cluster. Although these points are compared with the research index —cluster radius—again in the algorithm iteration, the excluded points are again included in the calculations if their distances from the stabilised cluster centroid are longer than the cluster radius. This can improve the clustering quality. Computerised tests as well as synthetic and real samples show that this method is able to improve the clustering quality by up to 41.85% in the best-case scenario. According to the findings, the proposed method is very beneficial to big data.