Typicality-based clustering
Marie-Jeanne Lesot
Int. Journal of Information Technology and Intelligent Computing Volume 1, Number 2, pp. 279-292, 2006.

Abstract

Typicality degrees are defined to build prototypes that characterise data subcategories, taking into account both the common points of the category members and their distinctive features as compared to other categories. In this paper, these principles are extended to the unsupervised learning framework, leading to a clustering algorithm robust to outliers that avoids overlapping areas between clusters and builds subgroups that are indeed both compact and separable. It does not require to use a Euclidean distance, which makes it possible to identify non-convex clusters.

[Edit]