Semi-Supervised Learning Based On Ensemble Algorithm

With the rapid development of computer technology and the Internet, the ability of data collection and storage has been improved greatly and plenty of data from science research and daily life has been gained. How to effectively use these data to analyze and mine its internal relations and useful information, with the purpose of improving working efficiency and the quality of life, is still a challenging problem. Thus, machine learning, data mining and pattern recognition are derived from the field of research.Semi-supervised learning and ensemble learning are two important machine learning paradigms, which both aim to improve the learning generalization performance. The difference between the two learning paradigms is that the former attempts to achieve strong generalization by exploiting unlabeled data, while the latter by using multiple learners. In this paper, we combine the advantages of semi-supervised learning and ensemble learning, where not only does it improve the learning performance by using the unlabeled samples but also it achieves strong generalization of the iterative learner by using the ensemble learning. This paper combines Bagging and AdaBoost algorithm to improve the Tri-Training where we propose the ENSSL algorithm, and then the experimental results indicate the efficiency of ENSSL algorithm. The main work is as follows:(1) Firstly we discuss the base classifier selection problem, and experiments evaluate the performance about the four base classifier algorithms. Based on different base classifier algorithms, experiments evaluate performance on three semi-supervised learning algorithms including Self-Training, Co-Training and Tri-Training. Also Experiments evaluate performance on four ensemble learning algorithms Bagging, AdaBoost, Vote and Stacking which combines different base classifier algorithms.(2) This paper focuses on Tri-Training algorithm. Because it makes use of simple resampling method to generate multiple classifiers, which leads to smaller between the classifiers. In addition, the final output doesn’t take the differences base classifiers into consideration. Aiming at these deficiencies, we further enhance the performance of Tri-Training, and ENSSL algorithm is proposed. The improved ENSSL algorithm combines the two commonly used ensemble leraning algorithms Bagging and AdaBoost, and the final output adopt the accuracy of weighted of the classifiers. Experimental results for inertial sensor-based human action recognition datasets from application of DLR indicate the efficiency of ENSSL algorithm, and experiments also show that the novel ENSSL algorithm is superior to the former three semi-supervised learning algorithm.