R/amrElm-ssl-train.R

In rladeira/amrElm: Affinity Matrix Regularized Extreme Learning Machine

#' Creates a model for AMR-ELM.
#'
#' @param l the number of hidden neurons
#' @param nl the number of labeled patterns
#' @param XTrain training data, numerical with zero mean and unit variance and patterns in the lines, attributes in the columns
#'        - the unlabeled patterns must came after the labeled ones
#' @param YTrain training data labels (only two classes, with labels equals to -1 or +1, and 0 for the unlabeled patterns)
#' @param affinity - only cosine implemented
#' @return The amrElm model for semissupervised problems - a list with:
#'          Z: hidden layer weights
#'          H: hidden layer output
#'          W: output layer weights
#'          affinity: the affinity used to generate the model (e.g.: cosine affinity)
#'          dataTrain: training data for generating affinity matrix.
#' @examples
#'
#' \dontrun{
#' library(amrElm)
#'
#' data(heart)
#'
#' data <- heart$data
#' labels <- heart$labels
#'
#' l <- 500
#' nl <- 50
#'
#' N <- nrow(data)
#' randomPatterns <- seq(N)
#'
#' data <- data[randomPatterns,]
#' labels <- labels[randomPatterns]
#'
#' nTrain <- floor(2*N/3)
#' nTest <- N - nTrain
#'
#' data <- data[randomPatterns,]
#' labels <- labels[randomPatterns]
#'
#' XTrain <- data[1:nTrain,]
#' XTest <- data[(nTrain+1):N,]
#'
#' YTrain <- labels[1:nTrain]
#' YTest[(nl+1):nTrain] <- 0
#'
#' model <- amrElmSSLTrain(l,nl,XTrain,YTrain)
#' testOutput <- amrElmTest(XTest, model)
#' }
#' @export
amrElmSSLTrain <- function(l, nl, XTrain, YTrain, affinity='cosine') {

  NL <- nl

  nAttrib <- dim(XTrain)[2]
  nTrain <- dim(XTrain)[1]

  # Get random Z matrix (hidden layer weight matrix)
  Z <- replicate(l, stats::runif(nAttrib+1,-0.5,0.5))

  # Add bias to the XTrain matrix
  Xaug <- as.matrix(cbind(replicate(nTrain,1),XTrain))


  # Apply the sigmoid function to the hidden layer's output matrix (H)
  H <- Xaug %*% Z
  H <- e1071::sigmoid(H)

  if (affinity=='cosine') {
    affinityMatrix <- cosine(XTrain)
  } else {
    stop('Wrong affinity option')
  }


  #Calculates the perturbated matrix (P %*% H) for AMRELM
  HL <-  affinityMatrix %*% H;

  #Calculates the output layer weight matrix (W) and the network output for the train data - AMELM
  W <- corpcor::pseudoinverse(HL[1:NL,]) %*% YTrain[1:NL];


  return(list(W=W,Z=Z,H=H,affinity='cosine',dataTrain=XTrain))
}