A clustering method associated pretopological concepts and k-means algorithm.

International audienceThe aim of this work is to define a clustering method starting from thepretopological results related to the minimal closed subset concepts which provideus the view of relations between groups in its structure; then, we consider this resultas the pre-treatment for some classical clustering algorithms. Especially, k-meansphilosophy is observed by its remarkable benefits. Thus we propose a new clusteringmethod in two processes such as structuring process and clustering one. Thismethod allows us to: obtain a data clustering for both of categorical and numericdata - exclude the limit in determination of cluster number a priori - and attainwell-shaped clusters whose shapes are not influenced on existence of outliers