In the past two decades, researchers have been focused on
finding meaningful subspaces such that the low dimensional
representation of high dimensional data can facilitate better classification
performance. For instance, principal component analysis
(PCA) [1] aims to find a subspace that preserves well the secondorder
statistics and captures maximal variability of the data, but
does not take into account class separation for the purpose of
classification. Discriminant analysis methods such as Fisher's linear
discriminant analysis (LDA) [2] and unsupervised discriminant
projection (UDP) [3], on the other hand, seek to obtain subspaces
where similarity criteria are enhanced, especially when the
Gaussianity assumption (as in the LDA) does not hold for the
training and testing data. Manifold approaches such as Isomap [4],
locally linear embedding (LLE) [5], and Laplacian eigenmap (LE) [6]