Ranking and empirical minimization of U-statistics

被引:222
作者
Clemencon, Stephan [1 ]
Lugosi, Gabor [2 ]
Vayatis, Nicolas [3 ]
机构
[1] Ecole Natl Super Telecommun Bretagne, Dept TSI Signal & Images, F-75014 Paris, France
[2] Univ Pompeu Fabra, Dept Econ & Business, Barcelona 08005, Spain
[3] Ecole Normale Super, Ctr Math & Leurs Applicat, F-94235 Cachan, France
关键词
statistical learning; theory of classification; VC classes; fast rates; convex risk minimization; moment inequalities; U-processes;
D O I
10.1214/009052607000000910
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
The problem of ranking/ordering instances, instead of simply classifying them, has recently gained much attention in machine learning. In this paper we formulate the ranking problem in a rigorous statistical framework. The goal is to learn a ranking rule for deciding, among two instances, which one is "better," with minimum ranking risk. Since the natural estimates of the risk are of the form of a U-statistic, results of the theory of U-processes are required for investigating the consistency of empirical risk minimizers. We establish, in particular, a tail inequality for degenerate U-processes, and apply it for showing that fast rates of convergence may be achieved under specific noise assumptions, just like in classification. Convex risk minimization methods are also studied.
引用
收藏
页码:844 / 874
页数:31
相关论文
共 47 条