Abstract：To improve the pattern recognition for highdimensional data, assuming that any two similar samples within the same class had similar sparse representations, a novel dimensionality reduction method of discriminative sparsity locality preserving projections (DSLPP) was proposed based on SPP and LPP. Through sparse learning and locality preserving projections, the good sparse representation was preserved by the proposed DSLPP， and the potential local manifold structure and the discrimination information of highdimensional data were also be well captured in the obtained subspace. The expression ability and the identifiability of high dimensional data were enhanced in the subspace. The experiments were completed on Yale, ORL and PIE29 face databases to compare DSLPP with PCA, LPP, NPE and SPP. The results show that the proposed DSLPP is an effective dimensionality reduction algorithm, and it can well improve the classification performance for highdimensional data in subspace.
2015, Vol. 36 
