On Transductive Support Vector Machines

Transductive support vector machines (TSVM) has been widely used as a means of treating partially labeled data in semisupervised learning. Around it, there has been mystery because of lack of understanding its foundation in generalization. This article aims to clarify several controversial aspects regarding TSVM. Two main results are established. First, TSVM performs no worse than its supervised counterpart SVM when tuning is performed, which is contrary to several studies indicating otherwise. The “alleged” inferior performance of TSVM is mainly because it was not tuned in the process, in addition to the involved minimization routines. Second, we utilize difference convex programming to derive a nonconvex minimization routine for TSVM, which compares favorably against some state-of-the-art methods. This, together with our learning theory lands some support to TSVM.