Local Rademacher complexities

被引:464
作者
Bartlett, PL
Bousquet, O
Mendelson, S
机构
[1] Univ Calif Berkeley, Dept Stat, Berkeley, CA 94720 USA
[2] Univ Calif Berkeley, Div Comp Sci, Berkeley, CA 94720 USA
[3] Max Planck Inst Biol Cybernet, Embirical Inference Dept, D-72076 Tubingen, Germany
[4] Australian Natl Univ, Ctr Math & Applicat, Inst Adv Studies, Canberra, ACT 0200, Australia
关键词
error bounds; Rademacher averages; data-dependent complexity; concentration inequalities;
D O I
10.1214/009053605000000282
中图分类号
O21 [概率论与数理统计]; C8 [统计学];
学科分类号
020208 ; 070103 ; 0714 ;
摘要
We propose new bounds on the error of learning algorithms in terms of a data-dependent notion of complexity. The estimates we establish give optimal rates and are based on a local and empirical version of Rademacher averages, in the sense that the Rademacher averages are computed from the data, on a subset of functions with small empirical error. We present some applications to classification and prediction with convex function classes, and with kernel classes in particular.
引用
收藏
页码:1497 / 1537
页数:41
相关论文
共 36 条
[1]  
Bartlett P. L., 2003, Journal of Machine Learning Research, V3, P463, DOI 10.1162/153244303321897690
[2]   Empirical minimization [J].
Bartlett, PL ;
Mendelson, S .
PROBABILITY THEORY AND RELATED FIELDS, 2006, 135 (03) :311-334
[3]   Model selection and error estimation [J].
Bartlett, PL ;
Boucheron, S ;
Lugosi, G .
MACHINE LEARNING, 2002, 48 (1-3) :85-113
[4]  
BARTLETT PL, 2005, IN PRESS J AM STAT A
[5]  
Boucheron S, 2003, ANN PROBAB, V31, P1583
[6]  
Boucheron S, 2000, RANDOM STRUCT ALGOR, V16, P277, DOI 10.1002/(SICI)1098-2418(200005)16:3<277::AID-RSA4>3.0.CO
[7]  
2-1
[8]  
Bousquet O, 2003, PROG PROBAB, V56, P213
[9]   A Bennett concentration inequality and its application to suprema of empirical processes [J].
Bousquet, O .
COMPTES RENDUS MATHEMATIQUE, 2002, 334 (06) :495-500
[10]  
BOUSQUET O, 2002, COMPUTATIONAL LEARNI, V2375, P59