Weighted Transductive Support Vector Machines for text classification

When labeled examples are limited, the training process of Transductive Support Vector Machines (TSVMs) incorporates a large number of unlabeled examples for text classification. This paper introduces Weighted Transductive Support Vector Machines (WTSVMs) to refine training. The concept of term weight is defined and a method to compute weight factor based on the degree of similarity between the test sample and labeled documents for unlabeled hypertext documents is presented. Then, these weight factors are used to reformulate the penalties on unlabeled samples in the transductive support vector machines. In the WTSVMs, test samples are treated discriminately according to their weight factors. Thus, the adjustment of the decision hyper-plane is refined, hence a more reliable decision function. This work also proposes an algorithm for training WTSVMs efficiently and reports some experimental results on benchmark problems.