kernlab

kernlab

Kernel-based machine learning.

Support Vector Machines

library(kernlab)

# Classification
model <- ksvm(Species ~ ., data = iris, kernel = "rbfdot")

# Regression
model <- ksvm(mpg ~ ., data = mtcars, type = "eps-svr")

# Predict
pred <- predict(model, newdata = test_df)

SVM Types

# Classification
"C-svc"      # C-classification (default)
"nu-svc"     # Nu-classification
"C-bsvc"     # Bound-constraint classification
"spoc-svc"   # Crammer-Singer multi-class
"kbb-svc"    # Weston-Watkins multi-class

# Regression
"eps-svr"    # Epsilon-SVR
"nu-svr"     # Nu-SVR
"eps-bsvr"   # Bound-constraint epsilon-SVR

# One-class
"one-svc"    # One-class SVM (novelty detection)

Kernels

# RBF (Gaussian) - default
model <- ksvm(y ~ ., data = df, kernel = "rbfdot")

# Linear
model <- ksvm(y ~ ., data = df, kernel = "vanilladot")

# Polynomial
model <- ksvm(y ~ ., data = df, kernel = "polydot")

# Sigmoid
model <- ksvm(y ~ ., data = df, kernel = "tanhdot")

# Laplacian
model <- ksvm(y ~ ., data = df, kernel = "laplacedot")

# Custom kernel parameters
rbf_kernel <- rbfdot(sigma = 0.1)
model <- ksvm(y ~ ., data = df, kernel = rbf_kernel)

Parameters

model <- ksvm(
  y ~ .,
  data = df,
  type = "C-svc",
  kernel = "rbfdot",
  kpar = list(sigma = 0.1),  # Kernel parameters
  C = 1,                      # Cost parameter
  epsilon = 0.1,              # Epsilon for SVR
  nu = 0.2,                   # Nu parameter
  cross = 5,                  # Cross-validation folds
  prob.model = TRUE,          # Enable probability estimates
  scaled = TRUE               # Scale features
)

Model Information

# Summary
model

# Support vectors
SVindex(model)
alpha(model)

# Cross-validation error
cross(model)

# Fitted values
fitted(model)

Probability Predictions

# Enable probabilities
model <- ksvm(y ~ ., data = df, prob.model = TRUE)

# Predict probabilities
pred_prob <- predict(model, newdata = test_df, type = "probabilities")