A Unified Variance Reduction-Based Framework for Nonconvex Low-Rank Matrix Recovery

Authors: Lingxiao Wang, Xiao Zhang, Quanquan Gu

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We further illustrate the superiority of our generic framework through several specific examples, both theoretically and experimentally. In this section, we present the experimental performance of our proposed algorithm for different models based on numerical simulations and real data experiments.
Researcher Affiliation Academia 1Department of Computer Science, University of Virginia, Charlottesville, Virginia, USA. Correspondence to: Quanquan Gu <qg5w@virginia.edu>.
Pseudocode Yes Algorithm 1 Low-Rank Stochastic Variance-Reduced Gradient Descent (LRSVRG) and Algorithm 2 Initialization are presented in Section 2.1.
Open Source Code No The paper does not provide an explicit statement or a link for the open-source code of the described methodology.
Open Datasets Yes In particular, we use three large recommendation datasets called Jester1, Jester2 and Jester3 (Goldberg et al., 2001), which contain anonymous ratings on 100 jokes from different users.
Dataset Splits No The paper mentions 'we randomly choose half of the ratings as our observed data, and predict the other half based on different matrix completion algorithms' and 'We perform 10 different observed/unobserved entry splittings', but does not explicitly specify standard train/validation/test splits with percentages or sample counts.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes For matrix completion, we consider the unknown low-rank matrix X in the following settings: (i) d1 = 100, d2 = 80, r = 2; (ii) d1 = 120, d2 = 100, r = 3; (iii) d1 = 140, d2 = 120, r = 4. ... (1) noisy case: the noise follows i.i.d. normal distribution with variance σ2 = 0.25 and (2) noiseless case. ... both algorithms use the same initialization method (Algorithm 2) with optimal parameters selected by cross validation.