A new semi-supervised algorithm using silhouette width for clustering arbitrary-scaled data

Abstract

In order to explore the information embedded in a given set of data, clustering has been successfully used to partition an unlabeled sample into meaningful classes. When we have any additional information regarding the original class-membership for a subset of the data either directly or indirectly, we should definitely use them in an attempt to carry out a better cluster analysis, which is known as semi-supervised clustering. This paper proposes such an algorithm, using the silhouette width, with the involvement of an appropriate distance measure that enforces its applicability to arbitrary-scaled measurements. It is shown to outperform the other well-established methods by the means of data study, with diverse applications including high-dimensional cases.