Abstract: Spectral clustering algorithm is an important one among clustering algorithms，and it uses the feature vectors of the similarity matrix calculated from the sample data set to cluster the sample data.However，the computational complexity and time consumption will increase markedly because of the large­scale calculation of eigen­decomposition when the spectral clustering is applied to large scale data sets.The use of sampling methods can effectively reduce the time consumed by the spectral clustering algorithm，but the relationship between the data subsets extracted by simple randomly sampling is too weak，which usually cannot reflect the distribution characteristics of the sample data sets accurately.Based on this and aimed at the computing characteristics of spectral clustering algorithm，a new sampling strategy different from the simple random sampling is designed and multiply used to generate multiple data subsets that can reflect the distribution characteristics of the sample data sets more accurately because of their coexisting relevance and otherness.Then each data subset is spectral clustered by NJW algorithm(the most classical spectral clustering algorithm，proposed by Ng A Y，Jordom M I and Weiss Y)and every clustering results can be mapped to the whole sample data set according to the nearest neighbor principle，generating a number of component clusters which both have relevance and otherness.Finally，the clustering results of the whole sample data set are integrated to get the final unified clustering partition.Experimental results show that applying the proposed sampling method to the spectral clustering algorithm is effective compared with the traditional NJW algorithm and efficient compared with the ensemble algorithm of spectral clustering based on simple sampling.