SGD Algorithms based on Incomplete U-statistics: Large-Scale Minimization of Empirical Risk - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Conference Papers Year : 2015

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

Abstract

In many learning problems, ranging from clustering to ranking through metric learning, empirical estimates of the risk functional consist of an average over tuples (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.
Fichier principal
Vignette du fichier
nips15_sgd.pdf (462.99 Ko) Télécharger le fichier
nips15_sgd_supp.pdf (398.08 Ko) Télécharger le fichier
Origin : Files produced by the author(s)
Origin : Files produced by the author(s)
Loading...

Dates and versions

hal-01214667 , version 1 (05-11-2015)

Identifiers

  • HAL Id : hal-01214667 , version 1

Cite

Guillaume Papa, Stéphan Clémençon, Aurélien Bellet. SGD Algorithms based on Incomplete U-statistics: Large-Scale Minimization of Empirical Risk. Annual Conference on Neural Information Processing Systems (NIPS), Dec 2015, Montréal, Canada. ⟨hal-01214667⟩
357 View
185 Download

Share

Gmail Facebook X LinkedIn More