Skip to Main content Skip to Navigation
Theses

Méthodes d'apprentissage statistique pour le ranking théorie, algorithmes et applications

Abstract : Multipartite ranking is a statistical learning problem that consists in ordering observations that belong to a high dimensional feature space in the same order as the labels, so that the observations with the highest label appear at the top on the list. This work aims to understand the probabilistic nature of the multipartite ranking problem in order to obtain theoretical guaranties for ranking algorithms. In that framework, the output of a ranking algorithm takes the form of a scoring function, a function that maps the space of the observations to the real line which order is induced using the values on the real line. The contributions of this manuscript are the following : first we focus on the characterization of the optimal solutions of multipartite ranking. A new condition on the likelihood ratios is introduced and shown to be necessary and sufficient to make the multipartite ranking well-posed. Then, we look at the criteria to assess the scoring function and propose to use a generalization of the ROC curve named the ROC surface. To be used in applications, the empirical counterpart of the ROC surface is studied and results on its consistency are stated. The second topic of research is the design of algorithms to produce scoring functions. The first procedure is based on the aggregation of scoring functions learnt from bipartite sub-problems. To the aim of aggregating the orders induced by the scoring function, we use a metric approach based on the Kendall- to find a median scoring function. The second procedure is a tree-based recursive method inspired by the TreeRank algorithm that can be viewed as a weighted version of CART. A simple modification is proposed to obtain an approximation of the optimal ROC surface using a piecewise constant scoring function. These procedures are compared to the state of the art algorithms for multipartite ranking using simulated and real data sets. The performances highlight the cases where our procedures are well-adapted, specifically when the dimension of the features space is much larger than the number of labels. Last but not least, we come back to the bipartite ranking problem in order to derive adaptive minimax rates of convergence. These rates are established for classes of distributions controlled by the complexity of the posterior distribution and a low noise condition. The procedure that achieves these rates is based on plug-in estimators of the posterior distribution and an aggregation using exponential weights.
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-00936092
Contributor : Sylvain Robbiano <>
Submitted on : Friday, January 24, 2014 - 2:47:39 PM
Last modification on : Friday, July 31, 2020 - 10:44:06 AM
Long-term archiving on: : Thursday, April 24, 2014 - 10:56:03 PM

Identifiers

  • HAL Id : tel-00936092, version 1

Collections

Citation

Sylvain Robbiano. Méthodes d'apprentissage statistique pour le ranking théorie, algorithmes et applications. Statistics [math.ST]. Telecom ParisTech, 2013. English. ⟨tel-00936092⟩

Share

Metrics

Record views

922

Files downloads

1340