(ICONIC, 2020-09-24) Gikera, Rufus; Mambo, Shadrack; Mwaura, Jonathan
Clustering is one of the most important tasks in exploratory data analysis [1, 55, 59]. K-means are the most popular clustering algorithms [51, 61]. This is because of their ability to adapt to new examples and to scale up to large datasets. They are also easily understandable and computationally faster [57, 60, 3, 62]. However, the number of clusters, K, has to be specified by the user [50]. Random process is the norm of searching for appropriate number of clusters, until convergence [53, 5]. Several variants of the k-means algorithm have been proposed, geared towards optimal selection of the K [8, 48]. The objective of this paper is to analyze the scaling up problems associated with these variants for optimizing K in the k-means clustering algorithms. Finally, a more enhanced hybrid autoencoder-based k-means will be developed and evaluated against the existing variants.