The LiblineaR.ACF package provides fast linear support vector machine training through coordinate descent with the Accelerated Coordinate Frequencies adaptation. It is a wrapper around a modified version of the LIBLINEAR C/C++ library for large-scale regularized linear classification. ACF is most useful for large data sets and large values of C.
The ACF feature is active only for the L2-regularized L1-loss and L2-loss support vector classification. Other types of the original LIBLINEAR package are not supported.
For more information on the C/C++ LIBLINEAR library itself, refer to:
R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin. LIBLINEAR: A Library for Large Linear Classification, Journal of Machine Learning Research 9(2008), 1871-1874.
Software available at http://www.csie.ntu.edu.tw/~cjlin/liblinear .
For more information on the ACF extension, refer to:
Tobias Glasmachers and Ürün Dogan. Accelerated Coordinate Descent with Adaptive Coordinate Frequencies. Asian Conference on Machine Learning (ACML), 2013, 72-86.
Software available at www.ini.rub.de/PEOPLE/glasmtbl/code/acf-cd .
devtools::install_github(repo="aydindemircioglu/LiblineaR.ACF") on the R command line,
make sure you have the latest devtools package installed.
To see ACF in action, download the news20 dataset from https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/. Then, the following code will load the data set and train a classifier with ACF.
library(SparseM) library(e1071) library(LiblineaR.ACF) c = 10000 s = read.matrix.csr("./news20.binary") m = LiblineaR.ACF(data = s$x, target = s$y, type = 1, cost = c, bias = FALSE)
All of this software is copyrighted by the list of authors in the DESCRIPTION file of the package and subject to the GNU GENERAL PUBLIC LICENSE, Version 2, see the file COPYING for details. The LIBLINEAR C/C++ code is copyright by the LIBLINEAR Project (Chih-Chung Chang, Chih-Jen Lin et al.)
The main insight of Accelerated Coordinate Frequencies is that especially on larger data sets not every data point is equally important. Thus, preference should be given to important points. The uniform random selection method of LIBLINEAR is replaced by sampling from a distribution that captures the importance of points, estimated online based on optimization progress. For more information refer to the paper 'Accelerated Coordinate Descent with Adaptive Coordinate Frequencies' by Tobias Glasmachers and Ürün Dogan.