Performance improvement of clustering method based on random swap algorithm

Abstract

This paper proposes a method for improving performance of clustering algorithm. Among unsupervised learning methods such as clustering, K-means has the advantage of being widely used and simple to implement, but has problems that is heavily influenced by initial centroids and may be got stuck at local minimum. To minimize these shortcomings, this work uses a clustering method with a random swap algorithm based on K-means++ that calculates the distance between data and the center point as a probability. The swap-based clustering method replaces the existing centroids and the concentrated centroids with each other during clustering steps, and then performing re-partitions, and fine-tuning steps with K-means++. The experimental results show the swap-based K-means++ clustering method proposed in this paper has better performance than other methods, as comparing the method proposed with the method of other clustering under various circumstances. As producing datasets of vehicles on general purpose for measuring performance in various experimental environments, we demonstrate the excellence of the proposed method.