A new method to improve the efficiency and accuracy of incremental singular value decomposition

Abstract

Singular value decomposition (SVD) has been widely used in machine learning. It lies at the root of data analysis, and it provides the mathematical basis for many data mining techniques. Recently, interest in incremental SVD has been on the rise because it is well suited to streaming data. In this paper, we propose a new algorithm of incremental SVD that is designed to improve both efficiency and accuracy during computation. More specifically, our proposed algorithm takes advantage of the special structures of arrowhead and diagonal-plus-rank-one matrices involved in updating SVD models to expedite the updating process. Moreover, because the singular values are computed independently, the proposed method can be easily parallelized. In addition, as this paper shows, increasing rank can lead to more accurate singular values in the updating process. Experimental results from synthetic and real data sets demonstrate gains in efficiency and accuracy in the updating process.