Low-rank sparse subspace for spectral clustering

• Main Authors: ZHU, Xiaofeng, ZHANG, Shichao, LI, Yonggang, ZHANG, Jilian, YANG, Lifeng, FANG, Yue
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2019
• Subjects: Affinity matrix; Clustering methods; Correlation; Feature extraction; Feature selection; Laplace equations; Redundancy; Sparse matrices; spectral clustering; subspace learning; Technological innovation; Computer Engineering; Databases and Information Systems
• Online Access: https://ink.library.smu.edu.sg/sis_research/4093; https://ink.library.smu.edu.sg/context/sis_research/article/5096/viewcontent/08417928.pdf
• Institution: Singapore Management University

The current two-step clustering methods separately learn the similarity matrix and conduct k means clustering. Moreover, the similarity matrix is learnt from the original data, which usually contain noise. As a consequence, these clustering methods cannot achieve good clustering results. To address these issues, this paper proposes a new graph clustering methods (namely Low-rank Sparse Subspace clustering (LSS)) to simultaneously learn the similarity matrix and conduct the clustering from the low-dimensional feature space of the original data. Specifically, the proposed LSS integrates the learning of similarity matrix of the original feature space, the learning of similarity matrix of the low-dimensional space, the transformation matrix for finding the low-dimensional feature space of the original data, and the low rank constraint making the similarity matrix of low-dimensional space to be the final clustering results, in a framework. Moreover, we propose an iterative optimization method to adaptively adjust each of the processes towards the goal of clustering performance, and thus enabling to output good clustering results. Extensive experiments were conducted on the synthetic datasets and benchmark datasets, and experimental results showed that our proposed LSS method achieved the best clustering performance in terms of two evaluation metrics (i.e., clustering ACCuracy (ACC) and Normalized Mutual Information (NMI)), compared to the state-of-the-art clustering methods.