关键词: 半监督学习, 构造性机器学习, co-training算法, tri-training算法, 覆盖算法

Abstract: Constructive Machine Learning（CML） algorithm needs larger number of labeled samples to train a classification network, but it is difficult to obtain a mass of labeled samples. So a constructive learning method based on co-training algorithm is designed. It divides the limited labeled samples into three equal training sets and uses CML algorithm to train three single classifiers respectively. The unlabeled samples are labeled by voting together through the three single classifiers to expand the three training sets in turn until they are not expanded. The expanded three training sets are united to generate the final classifier. Experiment is conducted on UCI data set and results show that the algorithm is more effective than the CML, Tri-CML, NB and Tri-NB algorithms.

Key words: semi-supervised, constructive machine learning, co-training algorithm, tri-training algorithm, covering