Online regularized learning with pairwise loss functions

被引:0
作者
Zheng-Chu Guo
Yiming Ying
Ding-Xuan Zhou
机构
[1] Zhejiang University,School of Mathematical Sciences
[2] State University of New York at Albany,Department of Mathematics and Statistics
[3] City University of Hong Kong,Department of Mathematics
来源
Advances in Computational Mathematics | 2017年 / 43卷
关键词
Pairwise learning; Online learning; Regularization; RKHS; 68Q32; 68T05; 62J02; 62L20;
D O I
暂无
中图分类号
学科分类号
摘要
Recently, there has been considerable work on analyzing learning algorithms with pairwise loss functions in the batch setting. There is relatively little theoretical work on analyzing their online algorithms, despite of their popularity in practice due to the scalability to big data. In this paper, we consider online learning algorithms with pairwise loss functions based on regularization schemes in reproducing kernel Hilbert spaces. In particular, we establish the convergence of the last iterate of the online algorithm under a very weak assumption on the step sizes and derive satisfactory convergence rates for polynomially decaying step sizes. Our technique uses Rademacher complexities which handle function classes associated with pairwise loss functions. Since pairwise learning involves pairs of examples, which are no longer i.i.d., standard techniques do not directly apply to such pairwise learning algorithms. Hence, our results are a non-trivial extension of those in the setting of univariate loss functions to the pairwise setting.
引用
收藏
页码:127 / 150
页数:23
相关论文
共 49 条
  • [1] Agarwal S(2009)Generalization bounds for ranking algorithms via algorithmic stability J. Mach. Learn. Res. 10 441-474
  • [2] Niyogi P(1950)Theory of reproducing kernels Trans. Am. Math. Soc. 68 337-404
  • [3] Aronszajn N(2005)Local Rademacher complexities Ann. Stat. 33 1497-1537
  • [4] Bartlett PL(2002)Rademacher and Gaussian complexities: risk bounds and structural results J. Mach. Learn. Res. 3 463-482
  • [5] Bousquet O(2016)Generalization bounds for metric and similarity learning Machine Learning Journal 102 115-132
  • [6] Mendelson S(2008)Improved risk tail bounds for online algorithms IEEE Trans. Inf. Theory 54 286-390
  • [7] Bartlett PL(2014)Learning performance of coefficient-based regularized ranking Neurocomputing 133 54-62
  • [8] Mendelson S(2010)Large scale online learning of image similarity through ranking J. Mach. Learn. Res. 11 1109-1135
  • [9] Cao Q(2008)Ranking and empirical minimization of U-statistics Ann. of Stat. 36 844-874
  • [10] Guo ZC(2014)Guaranteed classification via regularized similarity learning Neural Comput. 26 497-522