Inductive semi-supervised universum classification

The universum samples, which do not belong to any class of the current interest, have been proved helpful for semi-supervised classification (SSC). It actually opens up a new learning problem called semi-supervised universum classification (SSUC), which simultaneously utilizes the labeled, unlabeled and universum data for learning. However, the existing works for SSUC all utilize the universum samples in graph-based SSC methods, which belong to a branch for SSC based on the manifold assumption, and learn in the transductive style. In this paper, we concentrate on another important branch for SSC based on the cluster assumption, and attempt to utilize the universum samples for further enhancing the performance in the inductive learning style. Finally we develop a novel inductive SSUC method based on the cluster assumption. Empirical experiments on the USPS and MNIST datasets show the superiority of the proposed method compared with both the SSC and SSUC methods.