<!doctype html public "-//w3c//dtd html 4.0 transitional//en">
<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=iso-8859-1">
<meta name="description" content="logistic regression, support vector machines, linear classification, document classification">
<title>LIBLINEAR -- A Library for Large Linear Classification
</title>
</head>
<body text="#000000" bgcolor="#FFEFD5" link="#FF0000" vlink="#0000FF">
<h2>
LIBLINEAR -- A Library for Large Linear Classification
</h2>
<h3>
<a href="http://www.csie.ntu.edu.tw/~cjlin/mlgroup">Machine Learning Group</a> at
National Taiwan University
<br>
<a href=contributors.html>Contributors</a>
</h3>
<hr WIDTH="100%">
<img SRC="../images/new.gif" >
We recently released <a href=../libshorttext>LibShortText</a>,
a library for short-text classification and analysis.
It's built upon LIBLINEAR.
<p>
<img SRC="../images/new.gif" >
Version 1.94 released on November 12, 2013.
Following the recent change of LIBSVM, we slightly adjust the way class labels are handled
internally.
By default labels are ordered by their first occurrence in the training set.
Hence for a set with -1/+1 labels, if -1 appears first,
then internally -1 becomes +1. This has caused
confusion. Now for data with -1/+1 labels, we specifically
ensure that internally the binary SVM has positive data corresponding to the +1 instances. For developers, see changes in the subrouting group_classes of linear.cpp.
<p> An experimental version using 64-bit int is in
<a href=../libsvmtools>LIBSVM tools</a>. It
in theory can handle up to
2^64 instances/features if memory is enough.
<p>
We are interested in <b>large sparse</b> regression data.
Please let use know if you have some. Thank you.
<p>
<img SRC="../images/new.gif" >
<b>A practical guide to LIBLINEAR</b> is now available in the end of
<A HREF="../papers/liblinear.pdf">LIBLINEAR</a> paper.
<br>
<p>
Some extensions of LIBLINEAR are at LIBSVM Tools</a>.
<br>
<p>
LIBLINEAR is the winner of
<a href=http://largescale.first.fraunhofer.de/summary/>
ICML 2008 large-scale learning challenge</a>
(linear SVM track). It is also used for winning
KDD Cup 2010.
<hr>
<h3>
Introduction</h3>
<p><b>LIBLINEAR </b> is a <b>linear</b> classifier
for data with <b>millions</b> of
instances and features.
It
supports
<ul>
<li>
L2-regularized classifiers
<br>
L2-loss linear SVM,
L1-loss linear SVM, and
logistic regression (LR)
</li>
<li>
L1-regularized classifiers (after version 1.4)
<br>
L2-loss linear SVM and
logistic regression (LR)
</li>
<li>
L2-regularized support vector regression (after version 1.9)
<br>
L2-loss linear SVR and
L1-loss linear SVR.
</li>
</ul>
<p>
Main features of <b>LIBLINEAR</b> include
<ul>
<li> Same data format as
<a href=../libsvm><b>LIBSVM</b></a>, our general-purpose
SVM solver, and also similar usage
<li>Multi-class classification: 1) one-vs-the rest, 2)
Crammer & Singer</li>
<li>Cross validation for model selection</li>
<li>Probability estimates (logistic regression only)</li>
<li>Weights for unbalanced data</li>
<li>MATLAB/Octave, Java, Python, Ruby interfaces</li>
</ul>
<a href="#document">Documentation</a>
<p>
<a href=FAQ.html>FAQ is here</a>
<hr>
<h3>
When to use LIBLINEAR but not <a href=../libsvm>LIBSVM</a></h3>
There are some large data
for which with/without nonlinear
mappings gives similar performances.
<b>Without using kernels</b>, one can quickly train
a much larger set via a linear classifier.
<b>Document classification</b> is one such application.
In the following example (20,242 instances and 47,236 features; available on
<a href=../libsvmtools/datasets>LIBSVM data sets</a>), the cross-validation time is significantly
reduced by using LIBLINEAR:
<pre>
% time libsvm-2.85/svm-train -c 4 -t 0 -e 0.1 -m 800 -v 5 rcv1_train.binary
Cross Validation Accuracy = 96.8136%
345.569s
% time liblinear-1.21/train -c 4 -e 0.1 -v 5 rcv1_train.binary
Cross Validation Accuracy = 97.0161%
2.944s
</pre>
<b>Warning:</b>While LIBLINEAR's default solver
is very fast for document classification,
it may be <b>slow</b> in other situations.
See Appendix C of our
<a href=../papers/guide/guide.pdf>SVM guide</a>
about using other solvers in LIBLINEAR.
<p>
<b>Warning:</b>If you are a beginner and your
data sets are not large, you should
consider LIBSVM first.
<hr>
<h3>
<a NAME="download">
Download LIBLINEAR</h3>
The current release (Version 1.94, November 2013) of <b>LIBLINEAR </b>can be obtained by downloading
the
<a href="http://www.csie.ntu.edu.tw/~cjlin/cgi-bin/liblinear.cgi?+http://www.csie.ntu.edu.tw/~cjlin/liblinear+zip">
zip
file
</a>
or
<a href="http://www.csie.ntu.edu.tw/~cjlin/cgi-bin/liblinear.cgi?+http://www.csie.ntu.edu.tw/~cjlin/liblinear+tar.gz">
tar.gz
</a>
file.
<p>The package includes the source code in C/C++.
A README file with detailed explanation is
provided. For <b>MS Windows</b> users, there is a subdirectory in the zip
file containing binary executable files.
<p>Please read the <a href="COPYRIGHT">COPYRIGHT</a>
notice before using
<b>LIBLINEAR</b>.
<hr>
<h3>
<a NAME="document"></a>
Documentation and Codes used for experiments in our papers</h3>
<p>
R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and
<b>C.-J. Lin</b>.
<A HREF="../papers/liblinear.pdf">
LIBLINEAR: A library for large linear classification
</a>
<i><a href=http://www.jmlr.org>Journal
of Machine Learning Research</a></I>
9(2008), 1871-1874.
<p> The appendices of this paper give all <b>implementation details</b>
of LIBLINEAR.
<p> In the end of this paper there is a
<b>practical guide</b> to LIBLINEAR
<p>See also some examples in Appendix C of the
<a href=../papers/guide/guide.pdf>SVM guide</a>.
<p>
<a href="exp.html">Code used for experiments in our
LIBLINEAR
papers can be found here</a>.
<hr>
<h3>
Interfaces to LIBLINEAR</h3>
<center>
<table border=1>
<tr>
<th>Language</th>
<th>Description</th>
<th>Maintainers and Their Affiliation</th>
<th>Supported LIBLINEAR version</th>
<th>Link</th>
</tr>
<tr>
<td> <a NAME="matlab">MATLAB</td>
<td> A simple MATLAB interface</td>
<td>
LIBLINEAR authors at National
Taiwan University.
</td>
<td> The latest</td>
<td> <a href=#download>Included in LIBLINEAR package</a></td>
</tr>
<tr>
<td> <a NAME="octave">Octave</td>
<td> A simple Octave interface</td>
<td>
LIBLINEAR authors at National
Taiwan University.
</td>
<td> The latest</td>
<td> <a href=#download>Included in LIBLINEAR package</a></td>
</tr>
<tr>
<td> <a NAME="java">Java</td>
<td> Java version of LIBLINEAR</td>
<td>Benedikt Waldvogel
</td>
<td> 1.92</td>
<td> <a href="http://www.bwaldvogel.de/liblinear-java/">Java LIBLINEAR</a></td>
</tr>
<tr>
<td> <a name=python> Python </a></td>
<td> A python interface has been included in LIBLINEAR since version 1.6.</td>
<td> LIBLINEAR authors at National
Taiwan University.</td>
<td> The latest</td>
<td> <a href=#download>Included in LIBLINEAR package</a></td>
</tr>
<tr>
<td> <a NAME="java">Python</td>
<td> Python wrapper of LIBLINEAR</td>
<td> Uwe Schmitt
</td>
<td> 1.32</td>
<td> <a href="http://public.procoders.net/liblinear2scipy/src/dist/">Zip/tar.gz file</a></td>
</tr>
<tr>
<td> <a NAME="ruby">Ruby</td>
<td> A Ruby interface via SWIG</td>
<td>
Kei Tsuchiya (extended from the work of Tom Zeng)
</td>
<td> 1.93</td>
<td> <a href=https://github.com/kei500/liblinear-ruby>liblinear-ruby-swig</a></td>
</tr>
<tr>
<td> <a NAME="perl">Perl</td>
<td> A Perl interface</td>
<td>
Koichi Satoh
</td>
<td> 1.93</td>
<td> <a href=https://github.com/sekia/Algorithm-LibLinear>perl module</a></td>
</tr>
<tr>
<td> <a name=weka>Weka</a></td>
<td> Weka wrapper</td>
<td> Benedikt Waldvogel</td>
<td> 1.5</td>
<td>
<a href=http://wiki.pentaho.com/display/DATAMINING/LibLINEAR>Weka LIBLINEAR</a>
</td>
</tr>
<tr>
<td> <a name=R>R</a></td>
<td> R interface to LIBLINEAR</td>
<td> <a href=http://www.thibaulthelleputte.be>Thibault Helleputte</a></td>
<td> 1.8</td>
<td>
<a href=http://cran.r-project.org/web/packages/LiblineaR>R LIBLINEAR</a>
</td>
</tr>
<tr>
<td> <a name="lisp">Common LISP</a></td>
<td> Common Lisp wrapper of LIBLINEAR</td>
<td>
<a href="http://quotenil.com">Gábor Melis</a>
</td>
<td> 1.92</td>
<td>
<a href="http://quotenil.com/git/?p=cl-libsvm.git;a=summary">Common LISP wrapper</a>
</td>
</tr>
<tr>
<td> <a name=scilab>Scilab</a></td>
<td> </td>
<td>
Holger Nahrstaedt from the Technical University of Berlin
</td>
<td> 1.8</td>
<td>
<a href=http://atoms.scilab.org/toolboxes/libsvm>Scilab interface</a>
</td>
</tr>
</table>
</center>
<p>
<hr>
Please send comments and suggestions to <a href="../index.html">Chih-Jen
Lin</a>.
</body>
</html>