Classification Risk-Based Semi-supervised Ensemble Learning Algorithm

The existing semi-supervised ensemble learning algorithms commonly encounter the issue of information confusion in predicting unlabeled samples. To address this issue, a classification risk-based semi-supervised ensemble learning(CR-SSEL) algorithm is proposed. Classification risk is utilized as the criterion for evaluating the confidence of unlabeled samples. It can measure the degree of sample uncertainty effectively. By iteratively training classifiers and restrengthening the high confidence samples, the uncertainty of sample labeling is reduced and thus the classification performance of SSEL is enhanced. The impacts of learning parameters, training process convergence and improvement of generalization capability of CR-SSEL algorithm are verified on multiple standard datasets. The experimental results demonstrate that CR-SSEL algorithm presents the convergence trend of training process with an increase in the number of base classifiers and it achieves better classification accuracy. © 2024 Science Press. All rights reserved.