LIBSVM.jl

This is a Julia interface for LIBSVM and for the linear SVM model provided by LIBLINEAR.

Features:

• Supports all LIBSVM models: classification C-SVC, nu-SVC, regression: epsilon-SVR, nu-SVR and distribution estimation: one-class SVM
• Model objects are represented by Julia type SVM which gives you easy access to model features and can be saved e.g. as JLD file
• Supports ScikitLearn.jl API

Usage

LIBSVM API

This provides a lower level API similar to LIBSVM C-interface. See ?svmtrain for options.

using LIBSVM
using RDatasets
using Printf
using Statistics

# Load Fisher's classic iris data
iris = dataset("datasets", "iris")

# First four dimension of input data is features
X = Matrix(iris[:, 1:4])'

# LIBSVM handles multi-class data automatically using a one-against-one strategy
y = iris.Species

# Split the dataset into training set and testing set
Xtrain = X[:, 1:2:end]
Xtest  = X[:, 2:2:end]
ytrain = y[1:2:end]
ytest  = y[2:2:end]

# Train SVM on half of the data using default parameters. See documentation
# of svmtrain for options
model = svmtrain(Xtrain, ytrain)

# Test model on the other half of the data.
ŷ, decision_values = svmpredict(model, Xtest);

# Compute accuracy
@printf "Accuracy: %.2f%%\n" mean(ŷ .== ytest) * 100

Precomputed kernel

It is possible to use different kernels than those that are provided. In such a case, it is required to provide a matrix filled with precomputed kernel values.

For training, a symmetric matrix is expected:

K = [k(x_1, x_1)  k(x_1, x_2)  ...  k(x_1, x_l);
     k(x_2, x_1)
         ...                            ...
     k(x_l, x_1)        ...         k(x_l, x_l)]

where x_i is i-th training instance and l is the number of training instances.

To predict n instances, a matrix of shape (l, n) is expected:

KK = [k(x_1, t_1)  k(x_1, t_2)  ...  k(x_1, t_n);
      k(x_2, t_1)
          ...                            ...
      k(x_l, t_1)        ...         k(x_l, t_n)]

where t_i is i-th instance to be predicted.

Example

# Training data
X = [-2 -1 -1 1 1 2;
     -1 -1 -2 1 2 1]
y = [1, 1, 1, 2, 2, 2]

# Testing data
T = [-1 2 3;
     -1 2 2]

# Precomputed matrix for training (corresponds to linear kernel)
K = X' * X

model = svmtrain(K, y, kernel=Kernel.Precomputed)

# Precomputed matrix for prediction
KK = X' * T

ỹ, _ = svmpredict(model, KK)

ScikitLearn API

You can alternatively use ScikitLearn.jl API with same options as svmtrain:

using LIBSVM
using RDatasets

# Classification C-SVM
iris = dataset("datasets", "iris")
X = Matrix(iris[:, 1:4])
y = iris.Species

Xtrain = X[1:2:end, :]
Xtest  = X[2:2:end, :]
ytrain = y[1:2:end]
ytest  = y[2:2:end]

model = fit!(SVC(), Xtrain, ytrain)
ŷ = predict(model, Xtest)

# Epsilon-Regression

whiteside = RDatasets.dataset("MASS", "whiteside")
X = Matrix(whiteside[:, 3:3])  # the `Gas` column
y = whiteside.Temp

model = fit!(EpsilonSVR(cost = 10., gamma = 1.), X, y)
ŷ = predict(model, X)

MLJ API

The MLJ interface to LIBSVM.jl consists of the following models:

• classification: LinearSVC, SVC, NuSVC
• regression: EpsilonSVR, NuSVR
• outlier detection: OneClassSVM

Each model has a detailed document string, which includes examples of usage. Document strings can be accessed from MLJ without loading LIBSVM.jl (or its MLJ interface) as shown in the following example:

using MLJ     # or MLJModels 
doc("NuSVC", pkg="LIBSVM")

This assumes the version of MLJModels loaded is 0.15.5 or higher.

Credits

The LIBSVM.jl library is currently developed and maintained by Matti Pastell. It was originally developed by Simon Kornblith.

LIBSVM by Chih-Chung Chang and Chih-Jen Lin