Search Machine Learning Repository:
Revisiting the Nystrom method for improved large-scale machine learning
Authors: Alex Gittens and Michael Mahoney
Conference: Proceedings of the 30th International Conference on Machine Learning (ICML-13)
Abstract: We reconsider randomized algorithms for the low-rank approximation of SPSD matrices such as Laplacian and kernel matrices that arise in data analysis and machine learning applications. Our main results consist of an empirical evaluation of the performance quality and running time of sampling and projection methods on a diverse suite of SPSD matrices. Our results highlight complementary aspects of sampling versus projection methods, and they point to differences between uniform and nonuniform sampling methods based on leverage scores. We complement our empirical results with a suite of worst-case theoretical bounds for both random sampling and random projection methods. These bounds are qualitatively superior to existing bounds--- e.g., improved additive-error bounds for spectral and Frobenius norm error and relative-error bounds for trace norm error.
authors venues years
Suggest Changes to this paper.
Brought to you by the WUSTL Machine Learning Group. We have open faculty positions (tenured and tenure-track).