Online transductive support vector machines for classification

Abstract

Transductive support vector machine (TSVM) is one kind of transductive inference process, which combines labeled samples with unlabelled samples to derive the decision rules for classification tasks. Compared with the classical SVM, the transductive SVM is more robust and can achieve better performance. However, there are some disadvantages still being explored. One of the vital drawbacks is its computational costs. Usually, the SVM and TSVM models need to be retrained from scratch for parameter variations whenever any new samples become available. This problem has hindered the use of transduction learning in many real world applications. To resolve these problems, the online transductive support vector machine (OTSVM) is present to improve the generalization performance of the TSVM model. The OTSVM integrates the incremental learning/decremental unlearning to learn new incoming unlabeled samples one by one and modifies its model parameters at the same time. In addition to this, the OTSVM also dynamically adjusts the labels of unmatched samples. Simulation results illustrate that the OTSVM can improve classification performance and increase its computational efficiency.