Statistically Improving K-means Clustering Performance

Abstract

The traditional K-means algorithm is a cornerstone in unsupervised learning, providing a simple yet effective method for data clustering. However, its reliance on random initialization often leads to sub-optimal clustering results. This paper introduces an enhanced version of K-means algorithm, aimed at improving the clustering results. Our proposed methodology depends on identifying clusters that need to be partitioned and clusters that need to be merged through a series of statistical operation and iteratively resolve the problem leading to better clusters. We compare our proposed approach to K-Means and K-means++ algorithms on S-2, California housing prices, and EMNIST datasets showing performance improvements.