Input-Sparsity Low Rank Approximation in Schatten Norm

Authors: Yi Li, David Woodruff

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experiments The contribution of our work is primarily theoretical: an input sparsity time algorithm for low-rank approximation for any Schatten p-norm. In this section, nevertheless, we give an empirical verification of the advantage of our algorithm on both synthetic and real-world data. We focus on the most important case of the nuclear norm, i.e., p = 1. ... In Table 1 we report the median of ε1, the median of ε2, the median running time of both algorithms, among 50 independent runs for each k, and the median running time of a full SVD of A among 10 runs.
Researcher Affiliation Academia 1School of Physical and Mathematical Sciences, Nanyang Technological University 2Department of Computer Science, Carnegie Mellon University.
Pseudocode Yes Algorithm 1 Outline of the algorithm for finding a low-rank approximation
Open Source Code No The paper does not provide any links or statements about releasing the source code for their described methodology. It mentions using MATLAB but does not offer access to their implementation.
Open Datasets Yes KOS data. For real-world data, we use a word frequency dataset, named KOS, from UC Irvine.4 The matrix represents word frequencies in blogs and has dimension 3430 6906 with 353160 non-zero entries. (footnote 4 points to https://archive.ics.uci.edu/ml/datasets/Bag+of+Words)
Dataset Splits No The paper describes generating synthetic data and running the algorithm multiple times, but it does not specify explicit training, validation, or test dataset splits, nor does it mention cross-validation or predefined benchmark splits.
Hardware Specification Yes 3All tests are run under MATLAB 2019b on a machine of Intel Core i7-6550U CPU@2.20GHz with 2 cores.
Software Dependencies Yes 3All tests are run under MATLAB 2019b on a machine of Intel Core i7-6550U CPU@2.20GHz with 2 cores.
Experiment Setup Yes We adopt a simpler version of Algorithm 1 by taking S to be a COUNT-SKETCH matrix of target dimension k2 and both R and T to be identity matrices of appropriate dimension. For the Frobenius-norm solution, we also take S to be a COUNT-SKETCH matrix of target dimension k2. We choose n = 3000 and generate a random n n matrix A of independent entries, each of which is uniform in [0, 1] with probability 0.05 and 0 with probability 0.95. Since the regime of interest is k n, we vary k among {5, 10, 20}.