计算机应用 ›› 2012, Vol. 32 ›› Issue (03): 643-645.DOI: 10.3724/SP.J.1087.2012.00643

• 人工智能 • 上一篇    下一篇

结合全局和局部正则化的半监督二分类算法

吕佳1,2,3   

  1. 1.内蒙古大学 数学科学学院,呼和浩特 010021;
    2.重庆师范大学 计算机与信息科学学院,重庆 400047;
    3.中国农业大学 理学院,北京100083
  • 收稿日期:2011-08-17 修回日期:2011-11-16 发布日期:2012-03-01 出版日期:2012-03-01
  • 通讯作者: 吕佳
  • 作者简介:吕佳(1978-),女,四川达州人,副教授,博士研究生,主要研究方向:机器学习、最优化技术。
  • 基金资助:

    国家自然科学基金资助项目(10831009,10971223,11071252)。

Semi-supervised binary classification algorithm based on global and local regularization

Lü Jia1,2,3   

  1. 1. School of Mathematical Sciences, Inner Mongolia University, Hohhot Nei Mongol 010021, China;
    2.College of Computer and Information Science, Chongqing Normal University, Chongqing 400047, China;
    3.College of Science, China Agricultural University, Beijing 100083, China
  • Received:2011-08-17 Revised:2011-11-16 Online:2012-03-01 Published:2012-03-01

摘要: 针对在半监督分类问题中单独使用全局学习容易出现的在整个输入空间中较难获得一个优良的决策函数的问题,以及单独使用局部学习可在特定的局部区域内习得较好的决策函数的特点,提出了一种结合全局和局部正则化的半监督二分类算法。该算法综合全局正则项和局部正则项的优点,基于先验知识构建的全局正则项能平滑样本的类标号以避免局部正则项学习不充分的问题,通过基于局部邻域内样本信息构建的局部正则项使得每个样本的类标号具有理想的特性,从而构造出半监督二分类问题的目标函数。通过在标准二类数据集上的实验,结果表明所提出的算法其平均分类正确率和标准误差均优于基于拉普拉斯正则项方法、基于正则化拉普拉斯正则项方法和基于局部学习正则项方法。

关键词: 半监督学习, 二分类问题, 全局正则化, 局部正则化, 平滑

Abstract: As for semi-supervised classification problem, it is difficult to obtain a good classification function for the entire input space if global learning is used alone, while if local learning is utilized alone, a good classification function on some specified regions of the input space can be got. Accordingly, a new semi-supervised binary classification algorithm based on a mixed local and global regularization was presented in this paper. The algorithm integrated the benefits of global regularizer and local regularizer. Global regularizer was built to smooth the class labels of the data so as to lessen insufficient training of local regularizer, and based upon the neighboring region, local regularizer was constructed to make class label of each data have the desired property, thus the objective function of semi-supervised binary classification problem was constructed. Comparative semi-supervised binary classification experiments on some benchmark datasets validate that the average classification accuracy and the standard error of the proposed algorithm are obviously superior to other algorithms.

Key words: semi-supervised learning, binary classification problem, global regularization, local regularization, smooth

中图分类号: