Improved Local Search for Binary Matrix Factorization

Authors: Seyed Hamid Mirisaee, Eric Gaussier, Alexandre Termier

AAAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental With thorough experiments on both synthetic and real data, we demonstrate that the heuristic we propose significantly improves the quality of existing methods in a reasonable amount of time. We also show that our local search procedure performs well when compared to other heuristic strategies (Kanungo et al. 2002).
Researcher Affiliation Academia Seyed Hamid Mirisaee Univ. Grenoble Alps/CNRS Grenoble, France Hamid.Mirisaee@imag.fr Eric Gaussier Univ. Grenoble Alps/CNRS Grenoble, France Eric.Gaussier@imag.fr Alexandre Termier Univ. Rennes I/CNRS Rennes, France Alexandre.Termier@irisa.fr
Pseudocode No The paper describes algorithmic procedures but does not include structured pseudocode or clearly labeled algorithm blocks.
Open Source Code No For SVD and NMF, we used the Matlab built-in functions and for the rest we have used our own efficient implementations (all the codes are available from the authors upon request).
Open Datasets Yes Following the methodology used in (Uno, Kiyomi, and Arimura 2005), we examined both real world datasets and synthetic ones, all available at http://fimi.ua.ac.be/data/ (last visited 15-Nov-2014).
Dataset Splits No The paper does not provide specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) needed to reproduce the data partitioning.
Hardware Specification No The paper does not explicitly describe the hardware used to run its experiments, such as specific GPU/CPU models, processors, or cloud resources with specifications.
Software Dependencies No All the implementations have been done in Matlab which offers efficient matrix operations. For SVD and NMF, we used the Matlab built-in functions and for the rest we have used our own efficient implementations (all the codes are available from the authors upon request).
Experiment Setup Yes For projected NMF and SVD, we found the best projection points by applying a grid search, with a step size of 0.05. For efficiency reasons, we mainly focus on p = 1 in the p-opt local search. As mentioned before, we refer to the method proposed in this paper as 1-opt-BMF, to the standard implementation as 1-opt-Standard and to the 1-opt local search associated with UBQP as 1-opt-UBQP.