A Research On Transductive Support Vector Machine

Support vector machine(SVM),which was put forward by Vapnik, is a new kind ofmachine learning methodsbased on statistical learning theory. Originally, it was used tosolve a small sample problem. With the development of the statistical learning theory,SVM got a rapid revolution on the aspect of theory and application. SVM has improvedthe ability for high-dimensional data processing, especially since the kernel function isapplied. So, SVM has been researched and applied widely, such as text classification,medical diagnosis, image detection and digital authentication, etc.Based on supervised learning, the normal SVM classification method must betrained by a large number of samples, which need to be tagged manually. However, theprocess is not only inefficient but also very expensive. Therefore, many researchersapplied some semi-supervised method to the study of SVM, typically including: Bennettcombined the clustering assumption in semi-supervised learning and the standardizedforms of SVM and put forward a kind of semi-supervised support vector machine(S3VM); Under direct inference on unlabelled sample, Joachims got transductivesupport vector machine (TSVM)etc;This article makes a detailed research on TSVM. The algorithm of TSVM based onlabeling sample in pairs and dynamic adjustment method to solve the difficult problemof estimating the number of positive sample. In view of the imbalanced training samples,we put forward the fuzzy TSVM incremental algorithm and Semi-TSVM algorithm,based on different punished parameters for different unlabeled sample. Although theabove several kinds of improved algorithms make a better classification of accuracy, thetime of training also increases. For this situation, SLS-TSVM model avoids solving OPproblem in each iteration, it also uses area labeling principle to realize labeling theunlabeled sample. Through the simulation of real data, we can find SLS-TSVM not onlymains the accuracy of PTSVM, but also improves the efficiency of training.However, the research of SVM based on transductive learning is a direction of thedevelopment of the support vector machine. Although TSVM has got some progress, theresearch on its theory is still not perfect. The key point for further research of TSVM isto improve the performance of the classifier, through making full use of the distribution information of unlabelled samples.