Blitz: A Principled Meta-Algorithm for Scaling Sparse Optimization

Authors: Tyler Johnson, Carlos Guestrin

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

Reproducibility Variable Result LLM Response
Research Type Experimental In practice, our theoretical insights lead to very fast convergence times for ℓ1-regularized learning. In the sequential setting, BLITZ outperforms solvers such as GLMNET and LIBLINEAR, making BLITZ one of the fastest algorithms for high dimensional lasso and sparse logistic regression on a single machine. We then show additional gains for BLITZ in limited-memory and distributed regimes.
Researcher Affiliation Academia Tyler B. Johnson TBJOHNS@WASHINGTON.EDU Carlos Guestrin GUESTRIN@CS.WASHINGTON.EDU University of Washington, Seattle, WA 98195, USA
Pseudocode Yes Algorithm 1 Common Working Set Algorithm
Open Source Code No The paper describes its implementation details, such as 'we implement BLITZ in C++' and 'We implement each method in C++,' and mentions using the Rabit library (with its URL), but it does not provide an explicit statement or link for the release of its own BLITZ source code.
Open Datasets Yes Datasets are publicly available from LIBSVM4. URL: http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/. and the Criteo click-through rate dataset6. URL: http://labs.criteo.com/downloads.
Dataset Splits No The paper mentions using training data and evaluating performance on various datasets like FINANCE, RCV1, Webspam, and Criteo, but it does not provide specific details on how these datasets are partitioned into training, validation, and test sets, nor does it specify percentages or counts for these splits.
Hardware Specification Yes Our hardware is a 64-bit machine with 2.0 GHz Intel i7-2630QM processors, 8 GB memory, and 6 MB cache. and Our hardware is a 64-bit machine with 2.60 GHz Intel i5-4278U processors and a SATA HDD that achieves read rates of 100 MB/s.
Software Dependencies Yes GLMNET 1.9-8 (Friedman et al., 2010), LIBLINEAR 1.94 (Yuan et al., 2012), compiled with version 4.8.2 of the applicable GNU C/C++/Fortran compiler
Experiment Setup Yes We initialize BLITZ with a relatively small, easy subproblem (100 features in sequential setting and ϵ1 = 0.5). and We solve for 15 seconds using regularization λ = 0.05λMAX1 and a variety of fixed r (from (8)) and ϵ values... and We choose λ = 0.05λMAX to select a desirable number of features...