Abstract：To extract nonlinear patterns, preserve the manifold structure, and reduce the projection time, a Fast Kernel Supervised Locality Preserving Projection (FKSLPP) algorithm is proposed. This new algorithm firstly selects a subset of the training set by supervised cluster selection algorithm to do Subset Kernel Principal Component Analysis (SKPCA), and then Supervised Locality Preserving Projection (SLPP) is performed in SKPCA subspace. Experiments results show that compared with SLPP and some other popular feature extraction algorithms, FKSLPP can get higher recognition rates; compared with kernel projection algorithms of state of art, FKSLPP is much faster. In some datasets, FKSLPP can get same or higher recognition rates while costs only one-tenth processing time of the common kernel projection algorithms.
2011, Vol. 33 
