Gradient Hard Thresholding Pursuit for Sparsity-Constrained Optimization

Authors: Xiaotong Yuan, Ping Li, Tong Zhang

ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical evidences show that our method is superior to the state-of-the-art greedy selection methods when applied to learning tasks of sparse logistic regression and sparse support vector machines.
Researcher Affiliation Academia Xiao-Tong Yuan XTYUAN1980@GMAIL.COM Department of Statistical Science, Cornell University, Ithaca, NY 14853, USA Dept. of Statistics & Biostatistics, Dept. of Computer Science, Rutgers University, Piscataway, NJ 08854, USA Ping Li PINGLI@STAT.RUTGERS.EDU Dept. of Statistics & Biostatistics, Dept. of Computer Science, Rutgers University, Piscataway, NJ 08854, USA Tong Zhang TZHANG@STAT.RUTGERS.EDU Dept. of Statistics & Biostatistics, Rutgers University, Piscataway, NJ 08854, USA
Pseudocode Yes Algorithm 1: Gradient Hard Thresholding Pursuit (Gra HTP). Initialization: x(0) with x(0) 0 k (typically x(0) = 0), t = 1. Output: x(t). repeat (S1) Compute x(t) = x(t 1) η f(x(t 1)); (S2) Let F (t) = supp( x(t), k) be the indices of x(t) with the largest k absolute values; (S3) Compute x(t) = arg min{f(x), supp(x) F (t)}; t = t + 1; until halting condition holds; Fast Gra HTP repeat Compute x(t) = x(t 1) η f(x(t 1)); Compute x(t) = x(t) k as the truncation of x(t) with top k (in magnitude) entries preserved; t = t + 1; until halting condition holds;
Open Source Code No The paper does not include an unambiguous statement that the authors are releasing the code for the work described, nor does it provide a direct link to a source-code repository.
Open Datasets No The paper mentions specific datasets 'rcv1.binary' and 'news20.binary' but does not provide a direct URL, DOI, specific repository name, or a formal citation with author names and year for public access to these datasets.
Dataset Splits Yes For rcv1.binary, a training subset of size 20,242 and a testing subset of size 20,000 are used. For news20.binary, a training subset of size 10,000 and a testing subset of size 9,996 are used.
Hardware Specification Yes All the considered algorithms are implemented in Matlab 7.12 running on a desktop with Intel Core i7 3.2G CPU and 16G RAM.
Software Dependencies Yes All the considered algorithms are implemented in Matlab 7.12 running on a desktop with Intel Core i7 3.2G CPU and 16G RAM.
Experiment Setup Yes We fix the regularization parameter λ = 10 4 in the objective of (6). ... We simply initialize w(0) = 0 and set the stopping criterion as w(t) w(t 1) / w(t 1) 10 4.