A comparative analysis on prototype-based clustering methods

In the machine learning domain, clustering is a fundamental unsupervised learning operation which aims to partition the instances of a dataset into clusters (i.e, groups, subsets) such that instances within the same cluster are much similar to each other and much different from the other clusters. In the broad spectrum of clustering methods, prototype-based methods characterize each cluster through a prototype (i.e. centroid) and a relocation scheme is employed to iteratively redistribute the instances into the clusters, guided by an objective function. In this paper, several prototype-based methods are brought into focus, including K-means, K-medoids, K-medians, Fuzzy C-means and Kernel K-means. These algorithms are experimentally analyzed on several original benchmark datasets, distorted benchmark datasets and synthetic datasets. The comparative analysis is oriented in two main aspects: accuracy and sensitivity to noise and outliers.