Connected graph decomposition for spectral clustering

Abstract

This paper proposes a new spectral clustering method based on local Principal Component Analysis (PCA) and connected graph decomposition. Specifically, our method randomly select centroids of the data set to take global structure of data points into consideration, and then uses local PCA to preserve the local structure of data points for constructing the similarity matrix. Furthermore, our method employs the connected graph decomposition to partition the resulting similarity matrix to group data points into clusters. Experimental analysis on 12 UCI data sets showed that our proposed method outperformed the state-of-the-art clustering methods in terms of clustering performance.