We study the performance of empirical risk minimization (ERM), with respect to the quadratic risk, in the context of convex aggregation, in which one wants to construct a procedure whose risk is as close as possible to the best function in the convex hull of an arbitrary finite class . We show that ERM performed in the convex hull of is an optimal aggregation procedure for the convex aggregation problem. We also show that if this procedure is used for the problem of model selection aggregation, in which one wants to mimic the performance of the best function in itself, then its rate is the same as the one achieved for the convex aggregation problem, and thus is far from optimal. These results are obtained in deviation and are sharp up to logarithmic factors.
Nous étudions les performances de la procédure de minimisation du risque empirique, par rapport au risque quadratique, pour le problème d’agrégation convexe. Dans ce problème, on souhaite construire des procédures dont le risque est aussi proche que possible du risque du meilleur élément dans l’enveloppe convexe d’une classe finie de fonctions. Nous prouvons que la procédure obtenue par minimisation du risque empirique sur la coque convexe de est une procédure optimale pour le problème d’aggrégation convexe. Nous prouvons aussi que si cette procédure est utilisée pour le problème d’agrégation en sélection de modèle, pour lequel on souhaite imiter le meilleur dans , alors le résidu d’agrégation est le même que celui obtenue pour le problème d’agrégation convexe. Cette procédure est donc loin d’être optimale pour le problème d’agrégation en sélection de modèle. Ces résultats sont obtenus en déviation et sont optimaux à des facteurs logarithmiques prés.
Keywords: learning theory, aggregation theory, empirical process theory
@article{AIHPB_2013__49_1_288_0, author = {Lecu\'e, Guillaume and Mendelson, Shahar}, title = {On the optimality of the empirical risk minimization procedure for the convex aggregation problem}, journal = {Annales de l'I.H.P. Probabilit\'es et statistiques}, pages = {288--306}, publisher = {Gauthier-Villars}, volume = {49}, number = {1}, year = {2013}, doi = {10.1214/11-AIHP458}, mrnumber = {3060158}, zbl = {1259.62038}, language = {en}, url = {http://archive.numdam.org/articles/10.1214/11-AIHP458/} }
TY - JOUR AU - Lecué, Guillaume AU - Mendelson, Shahar TI - On the optimality of the empirical risk minimization procedure for the convex aggregation problem JO - Annales de l'I.H.P. Probabilités et statistiques PY - 2013 SP - 288 EP - 306 VL - 49 IS - 1 PB - Gauthier-Villars UR - http://archive.numdam.org/articles/10.1214/11-AIHP458/ DO - 10.1214/11-AIHP458 LA - en ID - AIHPB_2013__49_1_288_0 ER -
%0 Journal Article %A Lecué, Guillaume %A Mendelson, Shahar %T On the optimality of the empirical risk minimization procedure for the convex aggregation problem %J Annales de l'I.H.P. Probabilités et statistiques %D 2013 %P 288-306 %V 49 %N 1 %I Gauthier-Villars %U http://archive.numdam.org/articles/10.1214/11-AIHP458/ %R 10.1214/11-AIHP458 %G en %F AIHPB_2013__49_1_288_0
Lecué, Guillaume; Mendelson, Shahar. On the optimality of the empirical risk minimization procedure for the convex aggregation problem. Annales de l'I.H.P. Probabilités et statistiques, Volume 49 (2013) no. 1, pp. 288-306. doi : 10.1214/11-AIHP458. http://archive.numdam.org/articles/10.1214/11-AIHP458/
[1] Aggregated estimators and empirical complexity for least square regression. Ann. Inst. Henri Poincaré Probab. Stat. 40 (2004) 685-736. | EuDML | Numdam | MR | Zbl
.[2] Progressive mixture rules are deviation suboptimal. In Advances in Neural Information Processing Systems (NIPS), 2007.
.[3] Fast learning rates in statistical inference through aggregation. Ann. Statist. 37 (2009) 1591-1646. | MR | Zbl
.[4] Empirical minimization. Probab. Theory Related Fields 135 (2006) 311-334. | MR | Zbl
and .[5] -regularized linear regression: Persistence and oracle inequalities. Probab. Theory Related Fields 154 (2012) 193-224. | MR | Zbl
, and .[6] Model selection via testing: An alternative to (penalized) maximum likelihood estimators. Ann. Inst. Henri Poincaré Probab. Stat. 42 (2006) 273-325. | EuDML | Numdam | MR | Zbl
.[7] Some local measures of complexity of convex hulls and generalization bounds. In Computational Learning Theory (Sydney, 2002) 59-73. Lecture Notes in Comput. Sci. 2375. Springer, Berlin, 2002. | MR | Zbl
, and .[8] Sequential procedures for aggregating arbitrary estimators of a conditional mean. IEEE Trans. Inform. Theory 54 (2008) 1725-1735. | MR
and .[9] Aggregation for Gaussian regression. Ann. Statist. 35 (2007) 1674-1697. | MR | Zbl
, and .[10] Aggregation by exponential weighting and sharp oracle inequalities. In Learning Theory 97-111. Lecture Notes in Comput. Sci. 4539. Springer, Berlin, 2007. | MR | Zbl
and .[11] Lectures on Probability Theory and Statistics. Lecture Notes in Math. 1738. Springer, Berlin, 2000. | MR | Zbl
, and .[12] Functional aggregation for nonparametric regression. Ann. Statist. 28 (2000) 681-712. | MR | Zbl
and .[13] Minkowski symmetrizations suffice to arrive at an approximate Euclidean ball. Ann. of Math. (2) 156 (2002) 947-960. | MR | Zbl
.[14] Oracle Inequalities in Empirical Risk Minimization and Sparse Recovery Problems: Ecole d'Été de Probabilités de Saint-Flour XXXVIII-2008. Lecture Notes in Math. Springer, Berlin, 2011. | MR | Zbl
.[15] Local Rademacher complexities and oracle inequalities in risk minimization. Ann. Statist. 34 (2006) 2593-2656. | MR | Zbl
.[16] Aggregation via empirical risk minimization. Probab. Theory Related Fields 145 (2009) 591-613. | MR | Zbl
and .[17] General non-exact oracle inequalities in the unbounded case. Preprint.
and .[18] Probability in Banach Spaces: Isoperimetry and Processes. Ergebnisse der Mathematik und ihrer Grenzgebiete (3) [Results in Mathematics and Related Areas (3)] 23. Springer, Berlin, 1991. | MR | Zbl
and .[19] Obtaining fast error rates in nonconvex situations. J. Complexity 24 (2008) 380-397. | MR
.[20] Empirical processes with a bounded diameter. Geom. Funct. Anal. 20 (2010) 988-1027. | MR | Zbl
.[21] Reconstruction and sub-Gaussian operators in asymptotic geometric analysis. Geom. Funct. Anal. 17 (2007) 1248-1282. | MR | Zbl
, and .[22] Limit Theorems of Probability Theory. Oxford Studies in Probability 4. Oxford Univ. Press, New York, 1995. | MR | Zbl
.[23] The Volume of Convex Bodies and Banach Space Geometry. Cambridge Tracts in Math. 94. Cambridge Univ. Press, Cambridge, 1989. | MR | Zbl
.[24] Optimal rate of aggregation. In Computational Learning Theory and Kernel Machines (COLT-2003) 303-313. Lecture Notes in Artificial Intelligence 2777. Springer, Heidelberg, 2003. | Zbl
.[25] Optimal aggregation of classifiers in statistical learning. Ann. Statist. 32 (2004) 135-166. | MR | Zbl
.[26] Weak Convergence and Empirical Processes: With Applications to Statistics. Springer Ser. Statist. Springer, New York, 1996. | MR | Zbl
and .[27] Mixing strategies for density estimation. Ann. Statist. 28 (2000) 75-87. | MR | Zbl
.[28] Aggregating regression procedures to improve performance. Bernoulli 10 (2004) 25-47. | MR | Zbl
.Cited by Sources: