Adaptive Spectral Clustering Based on Grey Relational Analysis

As a method built upon spectral graph theory, spectral clustering has the advantages of processing data with any spatial shapes and converging on global optimal solutions. But it suffers from the defects that the clustering result is quite sensitive to its parameters and the number of clusters must be prespecified. In this paper, a novel approach which integrates the grey relational analysis based on difference information theory and a self-tuning method with spectral clustering is proposed. The similarities between data points are described by the balanced closeness degrees of their attribute sequences. A cost function is optimized to recognize the number of clusters automatically. So, the impact of the parameters can be eliminated and the performance can be improved. The experimental results proved the effectiveness of the new algorithm.