Safe transductive support vector machine

Abstract

Since semi-supervised learning can use fewer labelled samples to train a better model, semi-supervised methods are becoming popular in data mining. As an important algorithm of semi-supervised support vector machines (S (Formula presented.) VM), transductive support vector machine (TSVM) sometimes  may get worse models trained on both labelled samples and unlabelled samples than those trained only on labelled samples. To solve this problem, in this paper, we propose a safe TSVM (STSVM) based on the infinitesimal annealing algorithm. In the training of TSVM, we adopt the infinitesimal annealing and path following technology to approximate the step size of simulated annealing to balance the contradiction between annealing step and calculation time. During the annealing process, we call CP-step to update TSVM model with pseudo-labelled samples. If the current sample is on the boundary of combinatorial optimisation problem, SJ-step is called and a safety condition is designed to determine whether the sample needs to change its label or not, so as to ensure the TSVM model trained after changing is better than the model got before. The experimental results show that our STSVM algorithm can improve the accuracy of TSVM with a shorter running time, and is safer than the existing safe algorithms.