An automatic and efficient method for LAMOST’s massive spectral data reduction is presented in this paper. The identified cataclysmic variables were selected as templates to construct the feature space by PCA (the principal component analysis）, and most of the non-candidates were excluded by the method using support vector machine. Template matching strategy was used to identify the final candidates which were analyzed to complement the templates as feedback. Fifty eight new CVs candidates were found in the experiment, showing that our approach to finding special celestial bodies can be practical in LAMOST.