Titre :

SGD Algorithms based on Incomplete U -statistics: Large-Scale Minimization of Empirical Risk

Auteur(s) :

Papa, Guillaume [Auteur]
Laboratoire Traitement et Communication de l'Information [LTCI]
Clémençon, Stéphan [Auteur]
Département Images, Données, Signal [IDS]
Laboratoire Traitement et Communication de l'Information [LTCI]
Bellet, Aurelien [Auteur]
Machine Learning in Information Networks [MAGNET]

Titre de l’ouvrage :

SGD Algorithms based on Incomplete U -statistics: Large-Scale Minimization of Empirical Risk

Date de publication :

2015

Discipline(s) HAL :

Mathématiques [math]
Statistiques [stat]/Machine Learning [stat.ML]
Mathématiques [math]/Statistiques [math.ST]
Mathématiques [math]/Probabilités [math.PR]

Résumé en anglais : [en]

In many learning problems, ranging from clustering to ranking through metric learning, empirical estimates of the risk functional consist of an average over tu-ples (e.g., pairs or triplets) of observations, rather than ...
Lire la suite >In many learning problems, ranging from clustering to ranking through metric learning, empirical estimates of the risk functional consist of an average over tu-ples (e.g., pairs or triplets) of observations, rather than over individual observations. In this paper, we focus on how to best implement a stochastic approximation approach to solve such risk minimization problems. We argue that in the large-scale setting, gradient estimates should be obtained by sampling tuples of data points with replacement (incomplete U-statistics) instead of sampling data points without replacement (complete U-statistics based on subsamples). We develop a theoretical framework accounting for the substantial impact of this strategy on the generalization ability of the prediction model returned by the Stochastic Gradient Descent (SGD) algorithm. It reveals that the method we promote achieves a much better trade-off between statistical accuracy and computational cost. Beyond the rate bound analysis, experiments on AUC maximization and metric learning provide strong empirical evidence of the superiority of the proposed approach.Lire moins >

Langue :

Anglais

Audience :

Internationale

Vulgarisation :

Non

Collections :

• Centre de Recherche en Informatique, Signal et Automatique de Lille (CRIStAL) - UMR 9189

Source :

Harvested from HAL