svm_toolkit/README.rdoc

= SVM Toolkit 

source:: https://notabug.org/peterlane/svm_toolkit/

== Description

Support-vector machines are a popular tool in data mining.  This package
includes an amended version of the Java implementation of the libsvm library
(version 3.11).  Additional methods and examples are provided to support
standard training techniques, such as cross-validation, and simple
visualisations.  Training/testing of models can use a variety of built-in or
user-defined evaluation methods, including overall accuracy, geometric mean,
precision and recall.

== Features 

- All features of LibSVM 3.11 are supported, and many are augmented with Ruby wrappers.
- Loading Problem definitions from file in Svmlight, Csv or Arff (simple subset) format.
- Creating Problem definitions from values supplied programmatically in arrays.
- Rescaling of feature values.
- Integrated cost/gamma search for model with RBF kernel, taking advantage of multiple cores.
- Contour plot visualisation of cost/gamma search results.
- Model provides value of w-squared for hyperplane.
- svm-demo application, a version of the svm_toy applet which comes with libsvm.
- Model stores indices of training instances used as support vectors.
- User-selected evaluation techniques supported in Model#evaluate_dataset and
  Svm.cross_validation_search.
- Library provides evaluation classes for Cohen's Kappa statistics, F-measure,
  geometric-mean, Matthews Correlation Coefficient, overall-accuracy,
  precision, and recall.

== Example

The following example illustrates how a dataset can be constructed in code, and 
an SVM model created and tested against the different kernels. 

   require "svm_toolkit"
   include SvmToolkit
   
   puts "Classification with LIBSVM"
   puts "--------------------------"
    
   # Sample dataset: the 'Play Tennis' dataset 
   # from T. Mitchell, Machine Learning (1997)
   # --------------------------------------------
   # Labels for each instance in the training set
   #    1 = Play, 0 = Not
   Labels = [0, 0, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 0]
   
   # Recoding the attribute values into range [0, 1]
   Instances = [
     [0.0,1.0,1.0,0.0],
     [0.0,1.0,1.0,1.0],
     [0.5,1.0,1.0,0.0],
     [1.0,0.5,1.0,0.0],
     [1.0,0.0,0.0,0.0],
     [1.0,0.0,0.0,1.0],
     [0.5,0.0,0.0,1.0],
     [0.0,0.5,1.0,0.0],
     [0.0,0.0,0.0,0.0],
     [1.0,0.5,0.0,0.0],
     [0.0,0.5,0.0,1.0],
     [0.5,0.5,1.0,1.0],
     [0.5,1.0,0.0,0.0],
     [1.0,0.5,1.0,1.0]
   ]
   
   # create some arbitrary train/test split
   TrainingSet = Problem.from_array(Instances.slice(0, 10), Labels.slice(0, 10))
   TestSet     = Problem.from_array(Instances.slice(10, 4), Labels.slice(10, 4))
   
   # Iterate over each kernel type
   Parameter.kernels.each do |kernel|
    
     # -- train model for this kernel type
     params = Parameter.new(
       :svm_type => Parameter::C_SVC, 
       :kernel_type => kernel,
       :cost => 10, 
       :degree => 1,
       :gamma => 100
     )
     model = Svm.svm_train(TrainingSet, params)
   
     # -- test kernel performance on the training set
     errors = model.evaluate_dataset(TrainingSet, :print_results => true)
     puts "Kernel #{Parameter.kernel_name(kernel)} has #{errors} on the training set"
   
     # -- test kernel performance on the test set
     errors = model.evaluate_dataset(TestSet, :print_results => true)
     puts "Kernel #{Parameter.kernel_name(kernel)} has #{errors} on the test set"
   end

More examples can be found in the source code, linked above.

== Acknowledgements

The svm_toolkit is based on LibSVM, which is available from: 
http://www.csie.ntu.edu.tw/~cjlin/libsvm/

The contour plot uses the PlotPackage library, available from:
http://thehuwaldtfamily.org/java/Packages/Plot/PlotPackage.html