R/m_RidgeRegression.R

   1 #' @include b_LinearAlgorithm.R
   2
   3 #' @title Ridge Regression Algorithm
   4 #'
   5 #' @description Ridge Regression Algorithm.
   6 #' Inherits \code{\link{LinearAlgorithm}}
   7 #'
   8 #' @field lambda Value of lambda (let undefined for cross-validation). Default: undefined
   9 #' @field lambdas Vector of "optimal" lambda values over time. TODO: remove for production
  10 #'
  11 RidgeRegression = setRefClass(
  12     Class = "RidgeRegression",
  13
  14     fields = c(
  15         lambda = "numeric",
  16         lambdas = "numeric"
  17     ),
  18
  19     contains = "LinearAlgorithm",
  20
  21     methods = list(
  22         predict_noNA = function(XY, x)
  23         {
  24             if (length(lambda) > 0 || nrow(XY) < 30) #TODO: magic number
  25             {
  26                 #simple ridge regression with fixed lambda (not enough history for CV)
  27                 X = matricize(XY[,names(XY) != "Measure"])
  28                 Y = XY[,"Measure"]
  29                 lambda_ = ifelse(length(lambda) > 0, lambda, LAMBDA)
  30                 weight = ridgeSolve(X, Y, lambda_)
  31             }
  32
  33             else
  34             {
  35                 #enough data for cross-validations
  36                 require(MASS, quietly=TRUE)
  37                 gridLambda = seq(0.05,5.05,0.1)
  38                 res_lmr = lm.ridge(Measure ~ . + 0, data=XY, lambda = gridLambda)
  39                 lambda_ = res_lmr$lambda[which.min(res_lmr$GCV)]
  40                 weight = as.matrix(coef(res_lmr))[which.min(res_lmr$GCV),]
  41             }
  42
  43             lambdas <<- c(lambdas, lambda_)
  44
  45             appendWeight(weight)
  46             return (matricize(x) %*% weight)
  47         }
  48     )
  49 )