Liblinear-Ruby

Gem Version

Liblinear-Ruby is Ruby interface of LIBLINEAR using SWIG.
Now, this interface is supporting LIBLINEAR 2.1.

Installation

Add this line to your application's Gemfile:

gem 'liblinear-ruby'

And then execute:

$ bundle

Or install it yourself as:

$ gem install liblinear-ruby

Quick Start

This sample code execute classification with L2-regularized logistic regression.

require 'liblinear'

# train
model = Liblinear.train(
  { solver_type: Liblinear::L2R_LR },   # parameter
  [-1, -1, 1, 1],                       # labels (classes) of training data
  [[-2, -2], [-1, -1], [1, 1], [2, 2]], # training data
)
# predict
puts Liblinear.predict(model, [0.5, 0.5]) # predicted class will be 1

Parameter

There are some parameters you can specify:

  • solver_type
  • cost
  • sensitive_loss
  • epsilon
  • weight_labels and weights

solver_type

This parameter specifies a type of solver (default: Liblinear::L2R_L2LOSS_SVC_DUAL).
This corresponds to -s option on command line.
Solver types you can set are shown below:

# for multi-class classification
Liblinear::L2R_LR              # L2-regularized logistic regression (primal)
Liblinear::L2R_L2LOSS_SVC_DUAL # L2-regularized L2-loss support vector classification (dual)
Liblinear::L2R_L2LOSS_SVC      # L2-regularized L2-loss support vector classification (primal)
Liblinear::L2R_L1LOSS_SVC_DUAL # L2-regularized L1-loss support vector classification (dual)
Liblinear::MCSVM_CS            # support vector classification by Crammer and Singer
Liblinear::L1R_L2LOSS_SVC      # L1-regularized L2-loss support vector classification
Liblinear::L1R_LR              # L1-regularized logistic regression
Liblinear::L2R_LR_DUAL         # L2-regularized logistic regression (dual)

# for regression
Liblinear::L2R_L2LOSS_SVR      # L2-regularized L2-loss support vector regression (primal)
Liblinear::L2R_L2LOSS_SVR_DUAL # L2-regularized L2-loss support vector regression (dual)
Liblinear::L2R_L1LOSS_SVR_DUAL # L2-regularized L1-loss support vector regression (dual)

cost

This parameter specifies the cost of constraints violation (default 1.0).
This corresponds to -c option on command line.

sensitive_loss

This parameter specifies an epsilon in loss function of epsilon-SVR (default 0.1).
This corresponds to -p option on command line.

epsilon

This parameter specifies a tolerance of termination criterion.
This corresponds to -e option on command line.
The default value depends on a type of solver. See LIBLINEAR's README or Liblinear::Parameter.default_epsion for more details.

weight_labels and weights

These parameters are used to change the penalty for some classes (default []).
Each weights[i] corresponds to weight_labels[i], meaning that the penalty of class weight_labels[i] is scaled by a factor of weights[i].

Train

First, prepare training data.

# Define class of each training data:
labels = [1, -1, ...]

# Training data is Array of Array:
examples = [
  [1, 0, 0, 1, 0],
  [0, 0, 0, 1, 1],
  ...
]

# You can also use Array of Hash instead:
examples = [
  { 1 => 1, 4 => 1 },
  { 4 => 1, 5 => 1 },
  ...
]

Next, set the bias (this corresponds to -B option on command line):

bias = 0.5 # default -1

Then, specify parameters and execute Liblinear.train to get the instance of Liblinear::Model.

model = Liblinear.train(parameter, labels, examples, bias)

In this phase, you can save model as:

model.save(file_name)

If you have already had a model file, you can load it as:

model = Liblinear::Model.load(file_name)

Predict

Prepare the data you want to predict its class and call Liblinear.predict.

examples = [0, 0, 0, 1, 1]
Liblinear.predict(model, example)

Cross Validation

To get classes predicted by k-fold cross validation, use Liblinear.cross_validation.
For example, results[0] is a class predicted by examples excepts part including examples[0].

results = Liblinear.cross_validation(fold, parameter, labels, examples)

Thanks