Faster streaming algorithms for low-rank matrix approximations

Abstract

Low-rank matrix approximations are used in a significant number of applications. We present new algorithms for generating such approximations in a streaming fashion that expand upon recently discovered matrix sketching techniques. We test our approaches on real and synthetic data to explore runtime and accuracy performance. We apply our algorithms to the technique of Latent Semantic Indexing on a widely studied data set. We find our algorithms provide strong empirical results.