Cluster characterization through a representativity measure

Abstract

Clustering is an unsupervised learning task which provides a decomposition of a dataset into subgroups that summarize the initialbase and give information about its structure. We propose to enrich this result by a numerical coefficient that describes the cluster representativityand indicates the extent to which they are characteristic of the whole dataset. It is defined for a specific clustering algorithm, called OutlierPreserving Clustering Algorithm, opca, which detects clusters associated with major trends but also with marginal behaviors, in order tooffer a complete description of the inital dataset. The proposed representativity measure exploits the iterative process of opca to computethe typicality of each identified cluster.