Transferred Dimensionality Reduction

Dimensionality reduction is one of the widely used techniques for data analysis. However, it is often hard to get a demanded low-dimensional representation with only the unlabeled data, especially for the discriminative task. In this paper, we put forward a novel problem of Transferred Dimensionality Reduction, which is to do unsupervised discriminative dimensionality reduction with the help of related prior knowledge from other classes in the same type of concept. We propose an algorithm named Transferred Discriminative Analysis to tackle this problem. It uses clustering to generate class labels for the target unlabeled data, and use dimensionality reduction for them joint with prior labeled data to do subspace selection. This two steps run adaptively to find a better discriminative subspace, and get better clustering results simultaneously. The experimental results on both constrained and unconstrained face recognition demonstrate significant improvements of our algorithm over the state-of-the-art methods.