Université de Liège Réseau des Bibliothèques

BICTEL/e - ULg
Serveur institutionnel des thèses de doctorat



Nouvelles thèses
dans BICTEL/e - ULg
  • Palm, Mélanie - Etude de la faculté des protéines MX porcines à conférer une résistance contre le virus influenza A et contribution à l'identification du mécanisme impliqué
  • Dermine, Martin - TESTING CANDIDATE EFFECTORS CONTRIBUTING TO RESISTANCE TO PNEUMOVIRUSES
  • Cantinieaux, Dorothée - Le milieu conditionné de cellules souches mésenchymateuses: une alternative prometteuse à la transplantation de cellules souches mésenchymateuses après lésion de la moelle épinière
Présentation Recherche thèse Dépôt thèse Accès
gestionnaires
 
Page de résumé pour ULgetd-08252011-182524

Auteur : Meyer, Gilles
E-mail de l'auteur : G.Meyer@ulg.ac.be
URN : ULgetd-08252011-182524
Langue : Anglais/English
Titre : Geometric optimization algorithms for linear regression on fixed-rank matrices
Intitulé du diplôme : Doctorat en sciences de l'ingénieur
Département : FSA - Département d'électricité, électronique et informatique
Jury :
Nom : Titre :
Absil, Pierre-Antoine Membre du jury/Committee Member
Bach, Francis Membre du jury/Committee Member
Bonnabel, Silvère Membre du jury/Committee Member
Dhillon, Inderjit Membre du jury/Committee Member
Geurts, Pierre Membre du jury/Committee Member
Wehenkel, Louis Président du jury/Committee Chair
Sepulchre, Rodolphe Promoteur/Director
Mots-clés :
  • low-rank matrix
  • machine learning
  • optimization on manifolds
  • linear regression
Date de soutenance : 2011-09-02
Type d'accès : Public/Internet
Résumé :

Nowadays, large and rapidly evolving data sets are commonly encountered in many computer science applications. Efficiently mining and exploiting these data sets generally results in the extraction of valuable information and therefore appears as an important challenge in various domains including network security, computer vision, internet search engines, bioinformatics, marketing systems, online advertisement, social networks, just to name a few.

The rapid development of these modern computer science applications sustains an ever-increasing demand for efficient machine learning algorithms that can cope with large-scale problems, characterized by a large number of samples and a large number of variables.

The research reported in the present thesis is devoted to the design of efficient machine learning algorithms for large-scale problems. Specifically, we adopt a geometric optimization viewpoint to address the problem of linear regression in nonlinear and high-dimensional matrix search spaces. Our purpose is to efficiently exploit the geometric structure of the search space in the design of scalable linear regression algorithms.

Our search space of main interest will be the set of low-rank matrices. Learning a low-rank matrix is a typical approach to cope with high-dimensional problems. The low-rank constraint is expected to force the learning algorithm to capture a limited number of dominant factors that mostly influence the sought solution. We consider both the learning of a fixed-rank symmetric positive semidefinite matrix and of a fixed-rank non-symmetric matrix.

A first contribution of the thesis is to show that many modern machine learning problems can be formulated as linear regression problems on the set of fixed-rank matrices. For example, the learning of a low-rank distance, low-rank matrix completion and the learning on data pairs are cast into the considered linear regression framework. For these problems, the low-rank constraint is either part of the original problem formulation or is a sound approximation that significantly reduces the original problem size, resulting in a dramatic decrease in the computational complexity of algorithms.

Our main contribution is the development of novel efficient algorithms for learning a linear regression model parameterized by a fixed-rank matrix. The resulting algorithms preserve the underlying geometric structure of the problem, scale to high-dimensional problems, enjoy local convergence properties and confer a geometric basis to recent contributions on learning fixed-rank matrices. We thereby show that the considered geometric optimization framework offers a solid and versatile framework for the design of rank-constrained machine learning algorithms.

The efficiency of the proposed algorithms is illustrated on several machine learning applications. Numerical experiments suggest that the proposed algorithms compete favorably with the state-of-the-art in terms of achieved performance and required computational time.

Autre version :
Fichiers :
Nom du fichier Taille Temps de chargement évalué (HH:MI:SS)
Modem 56K ADSL
[Public/Internet] phd-thesis-meyer-color-links.pdf 1.70 Mb 00:04:03 00:00:09

Bien que le maximum ait été fait pour que les droits des ayants-droits soient respectés, si un de ceux-ci constatait qu'une oeuvre sur laquelle il a des droits a été utilisée dans BICTEL/e ULg sans son autorisation explicite, il est invité à prendre contact le plus rapidement possible avec la Direction du Réseau des Bibliothèques.


Parcourir BICTEL/e par Auteur|Département | Rechercher dans BICTEL/e


© Réseau des Bibliothèques de l'ULg, Grande traverse, 12 B37 4000 LIEGE