Data Segmentation with Improved K-Means Clustering Algorithm

Abstract

Unsupervised learning is also known as learning by observation in machine learning which groups the data instances based on their similarities. k-Means clustering technique is one of the most commonly used partition-based clustering methods that continuously relocate data instances from one cluster to another cluster to ameliorate the cluster validation. In this paper, we have introduced a new approach to improve the data clustering performance of the k-Means clustering algorithm. The proposed approach significantly reduces the number of iterations. Initially, we need to set the value of k, the number of clusters, and randomly select k number of instances from data as initial cluster centers. Then rest of the instances are assigned to the clusters based on the minimum Euclidean value. In the traditional k-means clustering method, each data instance is compared with each cluster center. But, in this proposed method we assign an instance into a cluster based on the average value of all instances that are already assigned to the cluster instead of the cluster center. The primary innovation lies in this modification of the assignment of instances into a cluster, which diverges significantly from conventional methodologies. By harnessing the in-place-mean of cluster instances calculation during assignments, the proposed approach significantly curtails the number of iterations required for convergence.