Hybrid attributes similarity measurement for spectral clustering

Similarity measurement for spectral clustering has been well-studied in recent years due to its crucial role on describing the intrinsic structure of data points. In this paper, we propose a hybrid attributes similarity measure method to process the Gaussian kernel affinity matrix. Compared with traditional global or local scale methods, our new similarity measurement has a rather robustness to reflect the multi-scale and complex structure dataset, and the affinity matrix is clearly block diagonal. Experiment results show that our algorithm can successfully obtain higher performance on both synthetic and real life dataset than the existing similarity measure methods.