Beyond L1: Faster and Better Sparse Models with skglm

Authors: Quentin Bertrand, Quentin Klopfenstein, Pierre-Antoine Bannier, Gauthier Gidel, Mathurin Massias

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide an extensive experimental comparison and we show state-of-the-art improvements on a wide range of convex and non-convex problems.
Researcher Affiliation Collaboration Quentin Bertrand Mila & Ude M, Canada quentin.bertrand@mila.quebec Quentin Klopfenstein Luxembourg Centre for Systems Biomedicine University of Luxembourg Esch-sur-Alzette, Luxembourg Pierre-Antoine Bannier Independent Researcher Gauthier Gidel Mila & Ude M, Canada CIFAR AI Chair Mathurin Massias Univ. Lyon, Inria, CNRS, ENS de Lyon, UCB Lyon 1, LIP UMR 5668, F-69342 Lyon, France
Pseudocode Yes Algorithm 1 skglm (proposed) input : X, β Rp, nout N, nin N, ws_size N, ϵ > 0
Open Source Code Yes We release skglm, a flexible, scikit-learn compatible package, which easily handles customized datafits and penalties.
Open Datasets Yes We use datasets from libsvm4 (Fan et al. 2008, see table 2).
Dataset Splits No The paper states 'Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Section 3.', however, Section 3 describes the benchmarking process and dataset usage but does not explicitly detail train/validation/test splits, percentages, or cross-validation methodology for the models themselves.
Hardware Specification No Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [N/A]
Software Dependencies No Our package relying on numpy and numba (Lam et al., 2015; Harris et al., 2020) is attached in the supplementary material. No version numbers are given for these dependencies.
Experiment Setup Yes skglm (Algorithm 1, ours), using M = 5 iterates for the Anderson extrapolation.