bmrm (version 3.3)

lpSVM: Linearly Programmed SVM

Description

Linearly Programmed L1-loss Linear Support Vector Machine with L1 regularization

Usage

svmLP(x, y, LAMBDA = 1, loss.weights = 1)
svmMulticlassLP(x, y, LAMBDA = 1, loss.weights = 1)
# S3 method for svmLP
predict(object, x, ...)

Arguments

x

a numeric data matrix to predict

y

a response factor for each row of x. It must be a 2 levels factor for svmLP, or a >=2 levels factor for svmMulticlassLP

LAMBDA

control the regularization strength in the optimization process. This is the value used as coefficient of the regularization term.

loss.weights

numeric vector of loss weights to incure for each instance of x. Vector length should match length(y), but values are cycled if not of identical size.

object

an object of class svmLP

...

unused, present to satisfy the generic predict() prototype

Value

the optimized weights matrix, with class svmLP

predict() return predictions for row of x, with an attribute "decision.value"

Functions

• svmLP: linear programm solving binary-SVM with L1-regularization and L1-norm

• svmMulticlassLP: linear programm solving multiclass-SVM with L1-regularization and L1-norm

Details

svmLP solves a linear program implementing a linear SVM with L1 regularization and L1 loss. It solves: min_w LAMBDA*|w| + sum_i(e_i); s.t. y_i * <w.x_i> >= 1-e_i; e_i >= 0 where |w| is the L1-norm of w

svmMulticlassLP solves a linear program implementing multiclass-SVM with L1 regularization and L1 loss. It solves: min_w LAMBDA*|w| + sum_i(e_i); s.t. <w.x_i> - <w.x_j> >= 1-e_i; e_i >= 0 where |w| is the L1-norm of w

Examples

# NOT RUN {
  x <- cbind(100,data.matrix(iris[1:4]))
  y <- iris$Species
  w <- svmMulticlassLP(x,y)
  table(predict(w,x),y)

  w <- svmLP(x,ifelse(y=="setosa","setosa","not_setosa"))
  table(predict(w,x),y)
# }