svmplus (version 1.0.1)

SVMP: Creates and returns an instance of the class specified in the svm_type.

Description

Creates and returns an instance of the class specified in the svm_type. In future, the current solver used for quadratic programming (quadprog) will be replaced by the equivaent quadprog solver defined in CVXR package. Also, LIBSVM and LIBLINEAR based faster implementaions are planned to be supported.

Usage

SVMP(cost = 1, gamma = 1, kernel_x = "rbf", degree_x = 3,
gamma_x = 0.001, kernel_xstar = "rbf", degree_xstar = 3,
gamma_xstar = 0.001, tol = 1e-05, svm_type = "QP")

Arguments

cost

cost of constraints violation

gamma

parameter needed for priviledged information

kernel_x

the kernel used for standard training data

degree_x

parameter needed for polynomial kernel for training data

gamma_x

parameter needed for rbf kernel for training data

kernel_xstar

the kernel used for priviledged information (PI)

degree_xstar

parameter needed for polynomial kernel for PI

gamma_xstar

parameter needed for rbf kernel for PI

tol

tolerance of dual variables

svm_type

optimization techiniques used: QP, LibSVM, LibLinear etc. Currently it supports only QP.

Value

an instance of the class specified in the svm_type. Currently it suports only "QP", hence returns instance of the class QPSvmPlus. The return instance can be used to call fit, project and predict methods of the QPSvmPlus.

Examples

# NOT RUN {
# This example is similar to the example given in the section 3.3 of the article:
# https://doi.org/10.1007/s10472-017-9541-2

#Generate train data
  mean1 = rep(0, 2)
  mean2 = rep(1, 2)
  cov2  = cov1 = .5 * diag(2)
  n = 20
  X1 = mvrnorm(n, mean1, Sigma = cov1)
  X2 = mvrnorm(n, mean2, Sigma = cov2)
  X_train = rbind(X1, X2)
  y_train = matrix(c(rep(1, n), rep(-1, n)), 2*n, 1)
# geberate privileged information data
  X1Star = matrix(0, n, 2)
  X2Star = matrix(0, n, 2)
  for(i in 1:n)
  {
    X1Star[i, 1] = norm(X1[i,] - mean1, type = "2")
    X1Star[i, 2] = norm(X2[i,] - mean2, type = "2")
  }
  for(i in 1:n)
  {
    X2Star[i, 1] = norm(X1[i, ] - mean2, type = "2")
    X2Star[i, 2] = norm(X2[i, ] - mean1, type = "2")
  }
  XStar = rbind(X1Star, X2Star)
# generate test data
  n_test = 10
  X1 = mvrnorm(n_test, mean1, Sigma = cov1)
  X2 = mvrnorm(n_test, mean2, Sigma = cov2)
  X_test = rbind(X1, X2)
  y_test = matrix(c(rep(1, n_test), rep(-1, n_test)), 2*n_test, 1)
# create instance of the class type QP, using RBF kernel
  qp = SVMP(cost = 100, gamma = .01,
            kernel_x = "rbf", gamma_x = .001,
            kernel_xstar = "rbf", gamma_xstar = .001,
            tol = .00001, svm_type = "QP")
# call the fit function
  qp$fit(X_train, XStar, y_train)
# call the predict function
  y_predict = qp$predict(X_test)
  print(length(y_predict[y_predict == y_test]))
  print("correct classification out of 20")



  #using polynomial kernel
  qp = SVMP(cost = 100, gamma = .01,
            kernel_x = "poly", degree_x = 3,
            kernel_xstar = "poly", gamma_xstar = 3,
            tol = .00001)

  qp$fit(X_train, XStar, y_train)

  y_predict = qp$predict(X_test)

  print(length(y_predict[y_predict == y_test]))
  print("correct classification out of 20")

  #using linear kernel
  qp = SVMP(cost = 10, gamma = .1,
            kernel_x = "linear",
            kernel_xstar = "linear",
            tol = .00001)

  qp$fit(X_train, XStar, y_train)

  y_predict = qp$predict(X_test)

  print(length(y_predict[y_predict == y_test]))
  print("correct classification out of 20")
# }