A Feature Space Alignment Learning Algorithm

Abstract

Manifold learning algorithms have been proved to be effective in pattern recognition, image analysis and data mining etc. The local tangent space alignment (LTSA) algorithm is a representative manifold learning method for dimension reduction. However, datasets cannot preserve their original features very well after dimension reduction by LTSA, due to the deficiency of local subspace construction in LTSA, especially when data dimensionality is large. To solve this problem, a novel subspace manifold learning algorithm called feature space alignment learning (FSAL in short) is proposed in this paper. In this algorithm, we employ candid covariance-free independent principal component analysis (CCIPCA) as a preprocessing step. Experiments over artificial and real datasets validate the effectiveness of the proposed method.