Research On Semi-supervised Learning Algorithms Based On Ensemble Learning

In pattern recognition domain, many issues acquire a large number of labeled datato train high-precision classifier, but labeled data is hard to obtain, even to consume alot of manpower and resources. With the development of data acquisition and storagetechnology, unlabeled data is easier to obtain than before. How to mining theinformation carried by unlabeled data and assisting a spot of labeled data forsemi-supervised learning, has became a research hotspot at home and abroad in recentyears.Based on the analysis of the existing semi-supervised learning algorithm, how totrain multiple classifiers to cooperate with each other for semi-supervised learningcombining with integrated learning techniques is a worthy research direction. This paperdoes research in this direction. the results achieved are as follows:(1) A semi-supervised learning algorithm named Vote-Training is proposed. A largenumber of experiments in the UCI data sets show that, Vote-Training algorithm can usethe higher recognition rate classifier trained by unlabeled samples effectively. With theexisting semi-supervised learning algorithm-Tri-Training compared, the algorithmconsumes less time, has more flexible structure, can adjust the voting strategy fordifferent problems and select the most appropriate way to address specific problems. Inthe experiments, the article does further analysis of the experimental data, points out theeffective preconditions of Vote-Training algorithm.(2)Analysis of the deficiency in traditional collaborative training algorithm,propose an improved collaborative training algorithm-CTA by adding more classifiersand introducing active learning techniques. Experiments on UCI data sets demonstratethe superiority of the algorithm.Semi-supervised learning algorithm has the problem of selecting superiority. Thereis no algorithm which can solve all of the semi-supervised learning problems. Analyzingthe existing algorithms further, using principles and techniques related, searching for aunified principle to guide the specific semi-supervised learning problem is a meaningfulresearch direction. This thesis is a good attempt.