ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2016, Vol. 53 ›› Issue (9): 1964-1970.doi: 10.7544/issn1000-1239.2016.20150436

• 人工智能 • 上一篇    下一篇

代价敏感大间隔分布学习机

周宇航,周志华   

  1. (计算机软件新技术国家重点实验室(南京大学) 南京 210023) (软件新技术与产业化协同创新中心(南京大学) 南京 210023) (zhouyh@lamda.nju.edu.cn)
  • 出版日期: 2016-09-01
  • 基金资助: 
    国家自然科学基金项目(61333014,61321491)

Cost-Sensitive Large Margin Distribution Machine

Zhou Yuhang, Zhou Zhihua   

  1. (National Key Laboratory for Novel Software Technology (Nanjing University), Nanjing 210023) (Collaborative Innovation Center of Novel Software Technology and Industrialization, Nanjing 210023)
  • Online: 2016-09-01

摘要: 在现实生活中的很多应用里,对不同类别的样本错误地分类往往会造成不同程度的损失,这些损失可以用非均衡代价来刻画.代价敏感学习的目标就是最小化总体代价.提出了一种新的代价敏感分类方法——代价敏感大间隔分布学习机(cost-sensitive large margin distribution machine, CS-LDM).与传统的大间隔学习方法试图最大化“最小间隔”不同,CS-LDM在最小化总体代价的同时致力于对“间隔分布”进行优化,并通过对偶坐标下降方法优化目标函数,以有效地进行代价敏感学习.实验结果表明,CS-LDM的性能显著优于代价敏感支持向量机CS-SVM,平均总体代价下降了24%.

关键词: 代价敏感学习, 间隔分布, 支持向量机, 表示定理, 对偶坐标下降法

Abstract: In many real world applications, different types of misclassification often suffer from different losses, which can be described by costs. Cost-sensitive learning tries to minimize the total cost rather than minimize the error rate. During the past few years, many efforts have been devoted to cost-sensitive learning. The basic strategy for cost-sensitive learning is rescaling, which tries to rebalance the classes so that the influence of different classes is proportional to their cost, and it has been realized in different ways such as assigning different weights to training examples, resampling the training examples, moving the decision thresholds, etc. Moreover, researchers integrated cost-sensitivity into some specific methods, and proposed alternative embedded approaches such as CS-SVM. In this paper, we propose the CS-LDM (cost-sensitive large margin distribution machine) approach to tackle cost-sensitive learning problems. Rather than maximize the minimum margin like traditional support vector machines, CS-LDM tries to optimize the margin distribution and efficiently solve the optimization objective by the dual coordinate descent method, to achieve better generalization performance. Experiments on a series of data sets and cost settings exhibit the impressive performance of CS-LDM; in particular, CS-LDM is able to reduce 24% more average total cost than CS-SVM.

Key words: cost-sensitive learning, margin distribution, support vector machine (SVM), representer theorem, dual coordinate descent method

中图分类号: