In this paper, a new sparse principal component analysis (SPCA) method, called DCPCA (sparse PCA using a difference convex program), is introduced as a spectral feature extraction technique in astronomical data processing. Using this method, we successfully derive the feature lines from the spectra of cataclysmic variables. We then apply this algorithm to get the first 11 sparse principal components and use the support vector machine (SVM) to classify. The results show that the proposed method is comparable with traditional methods such as PCA+SVM.
