BUFFONI David

Doutor em PhD
Equipe : MLIA
Data de partida : 31/07/2013
https://lip6.fr/David.Buffoni

Direção de pesquisa : Patrick GALLINARI

Co-supervisão£o : USUNIER Nicolas

Learning-to-Rank consistent surrogates for Information Retrieval tasks

In this era of technology, managing, controlling and retrieving informations sources has become a key part of our daily lives, and at the same time it presents a crucial challenge to researchers. In this thesis we tackle the problem of searching for items from a whole collection of objects, this is usually called Information Retrieval. We address in a Learning-to-Rank fashion where the goal is to learn a scoring function while minimizing a risk which reflects the quality of the ranked list. In practice, this risk cannot be directly optimized and the main goal is to design related surrogate losses thanks to the consistency property. Starting from this property, we show a way to derive two consistent surrogate loss functions with respect to some traditional Information Retrieval metrics. The resultant experiments prove our methodology.
All these theoretical considerations cannot be directly applied due to the fact that learning algorithms are sensitives to the data quality. We focus our attention on the preprocessing that needs to be done on the data to make Learning-to-Rank algorithms efficient in a two case study: XML retrieval and Text-Image Retrieval problems. For both, Learning-to-Rank algorithms are dependent on the quality of the supervision, the sampling of the training examples and the extracted features describing them. Finally, we conduct a series of experiments on these two problems to outperform traditional state-of-the-art Information Retrieval methods.

Defesas : 04/10/2012

Membros da banca :

Mohand Boughanem -- Université Paul Sabatier -- Rapporteur
Patrick Gallinari -- Université Pierre et Marie Curie -- Directeur
Patrice Perny -- Université Pierre et Marie Curie -- Examinateur
Liva Ralaivola -- Université Aix-Marseille -- Rapporteur
Nicolas Usunier -- Université Pierre et Marie Curie -- Co-directeur
Nicolas Vayatis -- Ecole Normale Supérieure de Cachan -- Examinateur

Data de partida : 31/07/2013

Publicações 2008-2015

Mentions légales
Mapa do site