Adaptively Discriminant Locality Preserving Projection

Abstract

Dimensionality reduction has been playing a significant role in many fields such as recognition, classification, clustering, high-dimensionality data compression. However, due to the existence of noises in the feature space of the original data, manifold learning methods take risks of finding the k nearest neighbors. LAPP designed a “coarse to fine” strategy to iteratively obtain the optimal subspace to solve this problem and obtain the optimal subspace. However, Since the discriminant information is also essential for the recognition and classification, ADLPP combined this “coarse to fine” idea with the idea of Supervised learning, which could not only preserve the local information after projection, solve the problem of noises and obtain the optimal subspaces, but also gain better performance on classification.