LAMOST-DR1 is the first data released by Guoshoujing telescop, which has the largest number of stellar spectra in the world at present. The data set provides the data source for searching for special and rare celestial objects like cataclysmic variable stars.Meanwhile, it requires more advanced astronomical data processing methods and techniques. A data mining method for cataclysmic variable spectra in massive spectra data is proposed in this paper. Different types of celestial spectra show obvious difference in the feature space constructed with Laplacian Eigenmap method. The parameters of artificial neural network are optimized with particle swarm optimization method and the total LAMOST-DR1 data is processed. 7 cataclysmic variable star spectra are found in the experiment including 2 dwarf nova, 2 nova like variables and a highly polarized AM Her type. The newly found spectra enrich the current cataclysmic variable spectra library. The experiment is the first attempt of searching for cataclysmic variable star spectra with Guoshoujing telescope data and the results show that our approach is feasible in LAMOST data. This method is also applicable for mining other special celestial objects in sky survey telescope data.