Skip to Main content Skip to Navigation

Classification paramétrique robuste partiellement supervisée en reconnaissance des formes

Abstract : Classifier design is a signicant stage of a pattern recognition process. One generally
distinguishes the supervised approach (classification) from the unsupervised one (clustering) according
to whether expertise on data is available or not. In this work, we study the intermediate
case of a semi-supervised clustering for mixed pool of numerical data. Whatever the approach,
some elements, called outliers, differ from the a priori model for the data and can therefore
disturb the clustering process. Robust clustering methods aim at limiting the influence of these
outliers either by modelling them explicitly, or by using robust estimators.
In the first part of this work, we study the concept of robustness through various algorithms
for clustering data. We focus on the use of so called M-estimators within the framework of estimation
based on likelihood maximization. The second part of this study deals with a state of
the art of semi-supervised clustering methods. We show that partial supervision is introduced by
modifying the objective function with a term of agreement with respect to membership degrees
or posterior probabilities fixed by the expert.
Finally, we propose a robust algorithm for clustering data in a partially supervised way. A reject
option is introduced. Classes are modelled by a mixture of two components whose parameters
are estimated through an iterative robust process. Rejection is achieved through assignment to
an additional class dedicated to outliers. The proposed approach have been successfully applied
on various artificial and real data sets.
Document type :
Complete list of metadatas

Cited literature [113 references]  Display  Hide  Download
Contributor : Christophe Saint-Jean <>
Submitted on : Saturday, May 12, 2007 - 11:20:50 AM
Last modification on : Wednesday, October 14, 2020 - 3:55:18 AM
Long-term archiving on: : Friday, September 21, 2012 - 2:46:22 PM


  • HAL Id : tel-00145895, version 1



Christophe Saint-Jean. Classification paramétrique robuste partiellement supervisée en reconnaissance des formes. Modélisation et simulation. Université de La Rochelle, 2001. Français. ⟨tel-00145895⟩



Record views


Files downloads