Scalable spectral ensemble clustering via building representative co-association matrix

Abstract

Ensemble clustering via building co-association matrix and combining multiple basic partitions from the same dataset into the consensus one has been widely used in spectral clustering and subspace clustering. However, with the ever-increasing cost of calculating the co-association matrix, the conventional ensemble clustering algorithm is no longer fit for dealing with the large-scale datasets due to its less scalability and time-consuming. In this paper, we propose a scalable spectral ensemble clustering method via building a representative co-association matrix to improve the ensemble clustering problem. Our method mainly includes constructing a sparse matrix to select the representative points and building the co-association matrix, and a robust and denoising representation for the co-association matrix can be learned through a low-rank constraint in a unified optimization framework. The experiments verify the high efficiency and scalability but less time cost of our method compared with state-of-art clustering methods in the six real-world datasets, especially in the large-scale datasets.