The Research On Collaboration-training Algorithm And Its Application

In the field of machine learning,the acquisition of huge quantities of unlabeled samples has become quite easy,but it is difficult to obtain labeled samples,even requiring professional personnel and equipment to label samples.Semi-supervised learning(SSL)effectively extracts useful information from unlabeled samples and makes it possible to build a good performance classifier with a small number of labeled samples,gradually SSL has become a hot research area in recent years at home and abroad.Collaboration-training,an important SSL branch,usually uses a small number of labeled samples to train two or more base classifiers,then each classifier can be trained further in the iterative learning process by using prediction results generated by other classifier(s),in turn each classifier's performance increases iteratively.However,when the number of labeled samples is small,the initial performance of the base classifier is often weak,which means wrong label is easily assigned to the unlabeled samples,sequentially classification accuracy is reduced.Aiming at base classifier weak initial performance problem,the thesis studies and improves collaboration-training algorithm by improving base classifier's initial performance and unlabeled sample selection strategy.The main contents of the thesis include:1.A collaboration-training algorithm for the problem,which the initial performance of the cooperative learning algorithm is weak,is proposed in the thesis.The algorithm is based on local and global consistency(LLGC)algorithm and can be divided into Co-LLSVM algorithm and Co-LLRF algorithm according to the differences of base classifier.The proposed algorithm makes best use of the LLGC capability that LLGC can accurately predict categories of unlabeled samples by using a small number of labeled samples.The proposed algorithm also provides abundant labeled samples for base classifier training and improves the classifier's initial accuracy,consequently increases the classifier performance.The experimental results with public data sets(7-sectors)show that the Co-LLSVM and Co-LLRF have better performance than the existing algorithm when the number of labeled samples is small,results also show Co-LLRF algorithm has better stability when dealing with sample imbalance problem and noise sample problem.2.Another collaboration-training algorithm is proposed and evaluated in order to improve the ability to select unlabeled samples in collaboration-training algorithm.The proposed algorithm combines confidence threshold handling method and voting confidence handling method in Tri-training to select unlabeled samples,thus the algorithm enhances the confidence level of the selected samples,ultimately the performance of the classifier is improved.The proposed algorithm is applied into side channel attack field,the experimental results on the DPA contest V4 dataset show that the proposed algorithm is more accurate than the existing collaboration-training algorithm or the supervised learning algorithm.The accuracy rate of proposed algorithm is significantly improved when only small amount of labeled power traces are available.