Comparison of three different spectral clustering methods based on graph segmentation

Based on the analysis of the realizing process of spectral clustering and the existing algorithms, three typical spectral clustering methods, such as Normalized cut, Min-max cut and automatic determination the number of clusters, were selected to discuss. The respective realization mechanism and clustering features were analyzed, and the experiment results of UCI (University of California, Irvine) data sets were compared. The research results show the clustering validity of the three algorithms and indicate the effect of threshold parameter and similarity measurement on performance of algorithms. Based on this, the feasible thoughts of spectral clustering to solve practical problems were introduced. That may be reference for engineering practice.