关键词: 谱聚类, 共享最近邻, 增量式, 大规模数据

Abstract:

The existing spectral clustering algorithm based on the closeness of shared nearest neighbors has been attracting more and more attentions, since it indeed well explores the potential similarity relationship between data points and has strong robustness for data sets that are not completely separated. However, it is still costly in the sense of both running time and memory requirements, which makes its clustering processing ability not very efficient. It has the disadvantages of slow running speed, long running time and invalid algorithm when facing large data sets, and hence becomes impractical for a large-scale data. In order to overcome these drawbacks, its incre-mental version is proposed in this paper. The basic idea is to first decompose the entire data set into its several subsets, and then the proposed spectral clustering algorithm guarantees its promising clustering performance by running on each subset in an incremental way. A lot of experiments on artificial data sets and simulation data sets indicate the effectiveness of the proposed spectral clustering algorithm. At the same time, the algorithm has low time consumption, high clustering accuracy, and can effectively cluster the increasing data sets.

Key words: spectral clustering, shared nearest neighbor, incremental, large-scale data