Clustering in Fisher Discriminative Subspaces

Charles Bouveyron (1, 2) , Camille Brunet (3)

1 SAMOS - Statistique Appliquée et MOdélisation Stochastique
2 CES - Centre d'économie de la Sorbonne
3 IBISC - Informatique, Biologie Intégrative et Systèmes Complexes

Clustering in high-dimensional spaces is nowadays a recurrent problem in many scientific domains but remains a difficult problem. This is mainly due to the fact that high-dimensional data usually live in low-dimensional subspaces hidden in the original space. This paper presents a model-based clustering approach which models the data in a discriminative subspace with an intrinsic dimension lower than the dimension of the original space. An estimation algorithm, called Fisher-EM algorithm, is proposed for estimating both the mixture parameters and the discriminative subspace. Experiments show that the proposed approach outperforms existing clustering methods and provides a useful representation of the high-dimensional data.

Mots clés: Model-based clustering; dimension reduction; discriminative subspaces; latent mixture model

Domaines: Théorie [stat.TH]; Statistiques [math.ST]; Méthodologie [stat.ME]; Machine Learning [stat.ML]

Charles Bouveyron, Camille Brunet. Clustering in Fisher Discriminative Subspaces. 13th International Conference on Applied Stochastic Models and Data Analysis (ASMDA 2009), Jun 2009, Vilnius, Lithuania. (elec. proc). ⟨hal-00375581⟩