Spectral energy minimization for semi-supervised learning

Abstract

Data mining problems often involve a large amount of unlabeled data and there is often very limited known information on the dataset. In such scenario, semi-supervised learning can often improve classification performance by utilizing unlabeled data for learning. In this paper, we proposed a novel approach to semi-supervised learning as as an optimization of both the classification energy and cluster compactness energy in the unlabeled dataset. The resulting integer programming problem is relaxed by a semi-definite relaxation where efficient solution can be obtained. Furthermore, the spectral graph methods provide improved energy minimization via the incorporation of additional criteria. Results on UCI datasets show promising results.