A new transductive learning method with universum data

Abstract

Transductive learning is a generalization of semi-supervised learning which attempts to learn a distinctive classifier from large amounts of unlabeled data. In addition, Universum data can bring prior knowledge to the classifier. The Universum data mean the data which do not belong to the positive or negative classes, which can improve the performance of the learning task. In this paper, we address the problem of transductive learning with Universum data, and propose a new method, called information entropy-based transductive support vector machine with Universum data(IEB-TUSVM), which mainly consists of two steps. In the first Nstep, we propose an information entropy-based method to select the informative examples from the source Universum data to obtain the informative Universum data. In the second step, we take the selected Universum data into the semi-supervised classification, which is further solved by the Lagrange method. We further analyze the computational complexity of the proposed method. Extensive experiments have shown that IEB-TUSVM method outperforms state-of-the-art semi-supervised methods and the Universum learning methods.