Research And Applications Of Sparse Affinity Spectral Clustering

Dense affinity is the key to increase the algorithm complexity. Although con-siderable progress has been made in research on sparse algorithms, these sparse methods applied to spectra analysis always cause low clustering quality and narrow application field of spectral clustering algorithm. The effectiveness of combining sparse method and spectral algorithm remains one of the greatest challenges at present. This paper devotes to a new algorithm based on the classical spectral clustering-Sparse Affinity Spectral Clustering(SASC),which takes advantage of the sparse similarity matrix in computation. SASC is widely expected to bring new hope for algorithm complexity. Further the clustering result of SASC is sensitive to the initial parameter values, so a clustering stability version of SASC is proposed-Sparse Spectral clustering based on K-Medoids(SSKM) taking the superiority of the PAM algorithm over KMeans. In addition, high dimensional data clustering is the focus of the current clustering analysis, but the two new algorithms, SASC and SSKM, are only suitable for low dimensional datasets, this paper therefore presents two high-dimensional versions of SASC and SSKM to ensure the utilization on high dimensional data, which investigate the High Correlation Filter (HCF) and the Prin-cipal Component Analysis (PCA) method to reduce or even eliminate the dimension disaster effects on high dimensional data processing. Under the different evaluation criteria, the new proposed algorithms have superiority in the simulation studies as well as real data.