LibSVM is an efficient solver for SVM classification and regression problems. The svm extension wraps this in a PHP interface for easy use in PHP scripts. The extension requires PHP 5.2 or above.
Libsvm itself is required, and is available through some package management, or direct from the website.
On ubuntu and other debian based systems, you may be able to installed the
apt-get install libsvm-dev
If installing from the website then some steps will need to be taken as the package does not install automatically:
wget http://www.csie.ntu.edu.tw/~cjlin/cgi-bin/libsvm.cgi?+http://www.csie.ntu.edu.tw/~cjlin/libsvm+tar.gz tar xvzf libsvm-3.1.tar.gz cd libsvm-3.1 make lib cp libsvm.so.1 /usr/lib ln -s libsvm.so.1 libsvm.so ldconfig ldconfig --print | grep libsvm
This last step should show libsvm is installed.
Once libsvm is installed, the extension can be installed in the usual way.
INSTALLING ON WINDOWS
A prebuilt win32 DLL is available from http://phpir.com/svm
The extension builds properly on Windows, if following the Windows step by step build process: https://wiki.php.net/internals/windows/stepbystepbuild
- Put SVM source in pecl/svm in php53dev directory
- Download latest libsvm, rename Makefile to Makefile.old and Makefile.win to Makefile - nmake
- Copy libsvm.lib from libsvm windows directory, and into deps/libs
- Copy svm.h into
buildconf configure --disable-all --enable-cli --with-svm=shared nmake
Make sure the libsvm.dll file from the libsvm windows directory in the path when running the resulting lib.
The basic process is to define parameters, supply training data to generate a model on, then make predictions based on the model. There are a default set of parameters that should get some results with most any input, so we'll start by looking at the data.
Data is supplied in either a file, a stream, or as an an array. If supplied in a file or a stream, it must contain one line per training example, which must be formatted as an integer class (usually 1 and -1) followed by a series of feature/value pairs, in increasing feature order. The features are integers, the values floats, usually scaled 0-1. For example:
-1 1:0.43 3:0.12 9284:0.2
In a document classification problem, say a spam checker, each line would represent a document. There would be two classes, -1 for spam, 1 for ham. Each feature would represent some word, and the value would represent that importance of that word to the document (perhaps the frequency count, with the total scaled to unit length). Features that were 0 (e.g. the word did not appear in the document at all) would simply not be included.
In array mode, the data must be passed as an array of arrays. Each sub-array must have the class as the first element, then key => value sets for the feature values pairs. E.g.
array( array(-1, 1 => 0.43, 3 => 0.12, 9284 => 0.2), );
This data is passed to the SVM class's train function, which will return an SVM model is successful.
$svm = new SVM(); $model = $svm->train($data);
or, for example with a filename
$svm = new SVM(); $model = $svm->train("traindata.txt");
Once a model has been generated, it can be used to make predictions about previously unseen data. This can be passed as an array to the model's predict function, in the same format as before, but without the label. The response will be the class.
$data = array(1 => 0.43, 3 => 0.12, 9284 => 0.2); $result = $model->predict($data);
In this case, $result would be -1.
Models can be saved and restored as required, using the save and load functions, which both take a file location.
And to load:
$model = new SVMModel(); $model->load('model.svm');