Journal of South China University of Technology (Natural Science Edition) ›› 2010, Vol. 38 ›› Issue (1): 87-91.doi: 10.3969/j.issn.1000-565X.2010.01.017

• Computer Science & Technology • Previous Articles     Next Articles

Semi-Supervised Classification Based on Regularization of Minimum Entropy

Liu Xiao-lan 1.2  Hao Zhi-feng3  Yang Xiao-wei2  Ma Xian-heng4   

  1. 1. School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, Guangdong, China; 2. School of Science, South China University of Technology, Guangzhou 510640, Guangdong, China; 3. Faculty of Computer, Guangdong University of Technology, Guangzhou 510090, Guangdong, China; 4. School of Software Engineering, South China University of Technology, Guangzhou 510006, Guangdong, China
  • Received:2008-12-30 Revised:2009-04-26 Online:2010-01-25 Published:2010-01-25
  • Contact: 刘小兰（1979-），女，博士生，讲师，主要从事人工智能、机器学习研究． E-mail:liuxl@scut．edu．cn
  • About author:刘小兰（1979-），女，博士生，讲师，主要从事人工智能、机器学习研究．
  • Supported by:

    广东省-教育部产学研结合项目（2007B090400031）;广东省科技计划项目（2008B080701005）

PDF

849

Abstract:

As the generative model needs modelling complex joint probability density and evaluating many parameters, a discriminant semi-supervised classification algorithm based on the regularization of minimum entropy is proposed. This algorithm uses Havrda-Charvat＇s structural α-entropy as the regularization item of the objective and employs the quasi-Newton method to solve the objective, which makes the algorithm discriminative and inductive and reduces the dependence of the algorithm on the model. At the same time, the algorithm can predict the labels of the out-of-sample data points easily. Simulated results of several UCI datasets demonstrate that the proposed algorithm is of low classification error even with few labeled data.

Key words: semi-supervised learning, conditional Havrda-Charvat＇s structural c~-entropy, regularization, pattern classification, quasi-Newton method