Co-regularization for classification

Abstract

Semi-supervised learning algorithms that combine labeled and unlabeled data receive significant interests in recent years and are successfully deployed in many practical data mining applications. Manifold regularization, one of the most representative works, tries to explore the geometry of the intrinsic data probability distribution by penalizing the classification function along the implicit manifold. Although existing manifold regularization, including Laplacian regularization (LR) and Hessian regularization (HR), yields significant benefits for partially labeled classification, it is observed that LR suffers from the poor generalization and HR exhibits the characteristic of instability, both manifold regularization could not accurately reflect the ground-truth. To remedy the problems in single manifold regularization and approximate the intrinsic manifold, we propose Manifold Regularized Co-Training (Co-Re) framework, which combines the manifold regularization (LR and HR) and the algorithm co-training. Extensive experiments on the USAA video dataset are conducted and validate the effectiveness of Co-Re by comparing it with baseline manifold regularization algorithms.