Méthodes de carte auto-organisatrice par mélange de lois contraintes. Application à l'exploration dans les tableaux de contingence textuels

Rodolphe Priam 1
1 TEXMEX - Multimedia content-based indexing
IRISA - Institut de Recherche en Informatique et Systèmes Aléatoires, Inria Rennes – Bretagne Atlantique
Abstract : This thesis is concerned with exploratory analysis of multidimensional data, which are often qualitative or textual, in particular Kohonen's self-organizing map models. The goal is to cluster and project simultaneously lines or columns of a data matrix. The result of these methods is a reduction in the form of a discrete surface of regression. We study more precisely mixture models of probabilistic laws: the parameters corresponding to means of clustered vectors are constrained by setting them at the nodes of a rectangular mesh. After an overview of these methods, and of the learning algorithms based on EM (Expectation - Maximization), we introduce two new approaches. The first one aims at generalizing the Correspondence Analysis method to large matrices: the CASOM algorithm is a naive Bayes classifier, which is constrained as a TPEM (Topology Preserving EM) for a contingency table. The second one consists in mutating image-clustering algorithms into map algorithms. As an illustration, we modify a clustering algorithm based on mean-field, and we get an algorithm named TNEM. We use these methods to ease the navigation in a textual corpus. Indeed, we provide objective criteria and cartographies.
Document type :
Theses
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-00532832
Contributor : Patrick Gros <>
Submitted on : Thursday, November 4, 2010 - 3:02:54 PM
Last modification on : Friday, November 16, 2018 - 1:23:53 AM
Long-term archiving on : Saturday, February 5, 2011 - 2:26:19 AM

Identifiers

  • HAL Id : tel-00532832, version 1

Citation

Rodolphe Priam. Méthodes de carte auto-organisatrice par mélange de lois contraintes. Application à l'exploration dans les tableaux de contingence textuels. Interface homme-machine [cs.HC]. Université Rennes 1, 2003. Français. ⟨tel-00532832⟩

Share

Metrics

Record views

377

Files downloads

542