Structured BFGS Method for Optimal Doubly Stochastic Matrix Approximation

Authors: Dejun Chu, Changshui Zhang, Shiliang Sun, Qing Tao

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We verify the advantages of our approach on both synthetic data and real data sets. The experimental results demonstrate that our algorithm outperforms the state-of-the-art solvers and enjoys outstanding scalability.
Researcher Affiliation Academia Dejun Chu1, Changshui Zhang2, Shiliang Sun3, Qing Tao4 1 School of Software, Hefei University of Technology 2 Department of Automation, Tsinghua University 3 School of Computer Science and Technology, East China Normal University 4 Army Academy of Artillery and Air Defense djun.chu@gmail.com, zcs@mail.tsinghua.edu.cn, slsun@cs.ecnu.edu.cn, taoqing@gmail.com
Pseudocode Yes Algorithm 1: Structured BFGS Algorithm for the Dual Problem (2) and Algorithm 2: Newton-based Line Search Method for Solving the Sub-problem (12)
Open Source Code Yes Our code will be released publicly on the github1 for reproducing all the results of this section. 1https://github.com/djchu/sbfgs4dsm
Open Datasets Yes We test 6 instances of the given matrices A which are derived from the real LIBSVM data sets at https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/.
Dataset Splits No The paper uses LIBSVM datasets and synthetic data, but does not explicitly describe specific train/validation/test dataset splits needed for reproduction.
Hardware Specification Yes All the algorithms have been implemented in MATLAB R2019b and run on a 3.00-GHz Intel Core i9 Linux machine with 128GB memory.
Software Dependencies Yes All the algorithms have been implemented in MATLAB R2019b
Experiment Setup Yes For the fairness of experimental comparison, we terminate three algorithms in the first two experiments with the same stopping tolerance, i.e., Fk ϵ = 10 12. ... We set σ = 1.0 in this subsection. ... We follow the setting of alternating projection algorithm (Zass and Shashua 2006) on six real data sets listed in Table 1.