A regularized approach for supervised multi-view multi-manifold learning from unlabeled data

摘要

In this paper, we combined two steps in a new multi-view multi-manifold learning algorithm that are essential for recognition tasks in the absence of class label information; first, we emphasize the first step of graph-based multi-view multi-manifold learning methods, i.e., select class-consistent neighbors from all available views. In multi-manifold space, the ideal neighborhood set is unidentified, and selection of a proper neighborhood set is not an easy task, especially if manifolds have some intersections. We describe each class of objects with continuous varying of pose angle as a relatively independent object-manifold. To find the object-manifolds, we utilize the transitivity of the similarity in the objects and use the TV-regularization to describe each object in a weighted sum of its class-consistent neighbors under different views. The proposed method aims to make a distinction between some objects with the same class and some objects with different classes that have similar views. The proposed method can be efficiently solved by an ADMM method. Second, we propose a regularized approach for supervised dimension reduction via discovering the discriminating information hidden in the data structure. Neighborhood selection and recognition accuracy experiments on COIL-20, CAS-PEAL, FEI, and ORL multi-view datasets have shown the excellent performance of our novel approach.