A co-training method based on entropy and multi-criteria

摘要

Co-training method is a branch of semi-supervised learning, which improves the performance of classifier through the complementary effect of two views. In co-training algorithm, the selection of unlabeled data often adopts the high confidence degree strategy. Obviously, the higher confidence of data signifies the higher accuracy of prediction. Unfortunately, high confidence selection strategy is not always effective in improving classifier performance. In this paper, a co-training method based on entropy and multi-criteria is proposed. Firstly, the data set is divided into two views with the same amount of information by entropy. Then, the clustering criterion and confidence criterion are adopted to select unlabeled data in view 1 and view 2, respectively. It can solve the problem that high confidence criterion is not always valid. Different choices can better play the complementary role of co-training, thus supplement what the other view does not have. In addition, the role of labeled data is fully considered in multi-criteria in order to select more valuable unlabeled data. Experimental results on several UCI data sets and one artificial data set show the effectiveness of the proposed algorithm.