A Method of PCA-SLPP Dimensionality Reduction for Feature Space Based on Manifold Learning

In view of the problem that the dimension of the feature space of complex pattern is so high that makes classification difficult, a dimensionality reduction method named PCA-SLPP was proposed based on manifold learning. Firstly, feature data of samples in the original feature set were analyzed with the method of principal component analysis (PCA) to make them uncorrelated. Then, with an improved supervised locality preserving projections (SLPP), data after PCA were mapped to make them more distinguishable while the manifold structure of feature data were preserved. The dimension of feature space was finally reduced according to the cumulative contribution in PCA and eigenvalue in SLPP. With the final dimensionality reduced data, a support vector machine was trained, and the macerals of inertinite of coal, which have complex structures, were classified. Experimental results show that, with PCA the redundancy of the feature data can be reduced effectively, which is helpful to improve the classification accuracy; when the dimension of PCA is constant and the total dimension is higher, the further reducing of dimension with the improved SLPP has fewer influence on the classification accuracy, but when total dimension is reduced to or below 2, the classification accuracy is decreased rapidly; with the proposed dimensionality reduction method, the classification accuracy is significantly higher than those from other algorithms when feature space dimension is reduced to or below half of the original dimension; the calculation time of the proposed method approaches to that of SLPP.