Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Global and Quadratic Convergence of Newton Hard-Thresholding Pursuit

Authors: Shenglong Zhou, Naihua Xiu, Hou-Duo Qi

JMLR 2021 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The efficiency of NHTP was demonstrated on both synthetic and real data in compressed sensing and sparse logistic regression. Keywords: sparse optimization, stationary point, Newton s method, hard thresholding, global convergence, quadratic convergence rate
Researcher Affiliation Academia Shenglong Zhou EMAIL School of Mathematical Sciences University of Southampton Southampton SO17 1BJ, UK Naihua Xiu EMAIL Department of Applied Mathematics Beijing Jiaotong University Beijing, China Hou-Duo Qi EMAIL School of Mathematical Sciences and CORMSIS University of Southampton Southampton SO17 1BJ, UK
Pseudocode Yes NHTP: Newton Hard-Thresholding Pursuit Step 0 Initialize x0. Choose η, γ > 0, σ (0, 1/2), β (0, 1). Set k 0. Step 1 Choose Tk T (xk; η). Step 2 If Tolη(xk; Tk) = 0, then stop. Otherwise, go to Step 3. Step 3 Compute the search direction dk by (22). Step 4 Find the smallest integer ℓ= 0, 1, . . . such that f(xk(βℓ)) f(xk) + σβℓ f(xk), dk . (27) Set αk = βℓ, xk+1 = xk(αk) and k k + 1, go to Step 1. Table 1: Framework of NHTP
Open Source Code No The paper provides links to code for benchmark methods (HTP, NIHT, GP, OMP, CoSaMP, SP, GraSP) but does not provide specific access information or a direct statement for the open-source release of the NHTP method described in this paper.
Open Datasets Yes Example 5 (Real data) This example comprises of seven real data sets for binary classification. They are colon-cancer1, arcene2, newsgroup3, news20.binary1, duke breastcancer1, leukemia1, rcv1.binary1, which are summarized in the following table, where the last three data sets have testing data. 1https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/ 2http://archive.ics.uci.edu/ml/index.php 3https://web.stanford.edu/~hastie/glmnet_matlab/
Dataset Splits Yes Example 5 (Real data) This example comprises of seven real data sets for binary classification. They are colon-cancer1, arcene2, newsgroup3, news20.binary1, duke breastcancer1, leukemia1, rcv1.binary1, which are summarized in the following table, where the last three data sets have testing data. Data name m samples n features training size m1 testing size m2 colon-cancer 62 2000 62 0 arcene 100 10000 100 0 newsgroup 11314 777811 11314 0 news20.binary 19996 1355191 19996 0 duke breast-cancer 42 7129 38 4 leukemia 72 7129 38 34 rcv1.binary 40242 47236 20242 20000
Hardware Specification Yes All experiments were conducted by using MATLAB (R2018a) on a desktop of 8GB memory and Inter(R) Core(TM) i5-4570 3.2Ghz CPU.
Software Dependencies Yes All experiments were conducted by using MATLAB (R2018a) on a desktop of 8GB memory and Inter(R) Core(TM) i5-4570 3.2Ghz CPU.
Experiment Setup Yes We initialize NHTP with x0 = 0 if f(0) = 0 and x0 = 1 if f(0) = 0. Parameters are set as σ = 10^-4/2, β = 0.5. For γ, theoretically any positive γ m2s is fine, but in practice to guarantee more steps using Newton directions, it is supposed to be relatively small (De Luca et al., 1996; Facchinei and Kanzow, 1997). Thus we choose γ = γk with updating ( 10^-10, if xk T c k = 0, 10^-4, if xk T c k = 0. For parameter η, in spite of that Theorem 10 has suggested to set 0 < η < η, it is still difficult to fix a proper one since M2s is not easy to compute in general. Overall, we choose to update η adaptively. Typically, we use the following rule: starting η with a fixed scalar associated with the dimensions of a problem and then update it as, η0 = 10(1 + s/n) min{10, ln(n)} > 1, ηk/1.05, if mod(k, 10) = 0 and Fηk(xk; Tk) > k^2, 1.05ηk, if mod(k, 10) = 0 and Fηk(xk; Tk) k^2, ηk, otherwise. where mod (k, 10) = 0 means k is a multiple of 10. We terminate our method if at kth step it meets one of the following conditions: Tolηk(xk; Tk) 10^-6, where Tolη(x, T) is defined as (15); |f(xk+1) f(xk)| < 10^-6(1 + |f(xk)|). k reaches the maximum number (e.g., 2000) of iterations.