An automatic and efficient method for cataclysmic variables candidates is presented in this paper. The nonlinear locally linear embedding-LLE method is applied in the newly released SDSS-DR8 spectra. Spectra are dimension-reduced by LLE and classified by artificial neural network. The greatly reduced final candidates can be identified manually. 6 new CVs candidates were found in the experiment, and the compare between LLE with PCA shows the feasibility of nonlinear method in data mining in astronomical data.