计算机工程与应用 ›› 2012, Vol. 48 ›› Issue (6): 155-157.

• 图形、图像、模式识别 • 上一篇    下一篇

一种利用Universum的半监督分类算法

杨 伟,侯臣平,吴 翊   

  1. 国防科学技术大学 数学与系统科学系,长沙 410073
  • 收稿日期:1900-01-01 修回日期:1900-01-01 出版日期:2012-02-21 发布日期:2012-02-21

Semi-supervised classification with Universum

YANG Wei, HOU Chenping, WU Yi   

  1. Department of Mathematics and System Science, National University of Defense Technology, Changsha 410073, China
  • Received:1900-01-01 Revised:1900-01-01 Online:2012-02-21 Published:2012-02-21

摘要: 分类是机器学习领域的重要分支,利用少量的标签数据进行分类和高维数据的分类是近期研究的热点问题。传统的半监督方法能够有效利用标签样本数据或非标签样本数据,但忽略了相关的非样本数据,即Universum。利用Universum的半监督分类算法,基于线性回归和子空间学习模型,结合了传统半监督方法和利用Universum方法两者的优点,在不增加标签数据的条件下显著地提高了高维数据的分类效果。仿真实验和真实数据上的分类结果都验证了算法的有效性。

关键词: 半监督分类, Universum方法, 线性回归, 子空间学习

Abstract: Classification is an important branch of machine learning. It remains a hot issue how to attain a better classification with less labeled data in recent research. Traditional semi-supervised classification can take advantage of the training samples, either labeled or unlabeled, but ignores related non-samples, called the Universum. Combining the advantage of traditional semi-supervised methods and the Universum, Semi-Supervised Classification with the Universum(SSCU) via linear regression and subspace learning, can effectively improve the classification of original high-dimensional data adding no labels. The effectiveness is verified by both simulation and real-world data.

Key words: semi-supervised classification, Universum, linear regression, subspace learning