R/Metafeatures.R

In nunompmoniz/autoresampling: Automated Resampling Strategies for Imbalanced Learning

#' Meta-Feature Description
#'
#' Provides anextensive description of a data set based on the extraction of meta-features
#'
#' @param ds A data set
#' @param form A model formula
#' @param numCores number of cores for parallel computing
#'
#' @return A vector of meta-features describing the data set
#' @export
#'
#' @examples
#' \dontrun{
#'
#' library(mlbench)
#'
#' data(PimaIndiansDiabetes)
#'
#' form <- diabetes ~ .
#'
#' getMetaFeatures(PimaIndiansDiabetes,form)
#'
#' }
#'
getMetaFeatures <- function(ds,form,numCores=1) {

  tgt <- which(colnames(ds)==as.character(form[[2]]))
  nms <- classNames(form, ds)

  # Number of cases in the dataset
  ncases <- nrow(ds)

  # Imbalance Ratio
  imbRatio <- nrow(ds[ds[,tgt]==nms[1],]) / nrow(ds[ds[,tgt]==nms[2],])

  # Number of Attributes
  nattributes <- ncol(ds)-1

  # Number of numeric and nominal variables
  numvars <- length(as.numeric(which(sapply(ds[,-tgt],is.numeric))))
  nomvars <- nattributes - numvars

  # Ratio of cases per attributes
  ratioCasesAttr <- ncases/nattributes

  # Fraction of numerical attributes showing outliers
  attributesWithOutliers <- attrWithOutliers(ds,tgt) / nattributes

  # IQR
  statsIQR <- statsIQR_NumAttrs(ds,tgt)

  # Coefficient of Variation
  statsCoefVar <- statsCoV_NumAttrs(ds,tgt)

  # Correlation between Numerical Attributes
  statsCorNumAttrs <- statsCor_NumAttrs(ds,tgt)

  # Geary's Kurtosis of Numerical Attributes
  statsGKurNumAttrs <- statsGKur_NumAttrs(ds,tgt)

  # Pearson's Kurtosis of Numerical Attributes
  statsPKurNumAttrs <- statsPKur_NumAttrs(ds,tgt)

  # Skewness of Numerical Attributes
  statsSkewNumAttrs <- statsSkew_NumAttrs(ds,tgt)

  # Do a single call to Minerva
  tmp <- statsALL_NumAttrs(ds,tgt,numCores=numCores)

  # Maximal Information Coefficient (MIC)
  #statsMICNumAttrs <- statsMIC_NumAttrs(ds,tgt)
  statsMICNumAttrs <- tmp$MIC
  # Maximum Asymmetry Score (MAS)
  #statsMASNumAttrs <- statsMAS_NumAttrs(ds,tgt)
  statsMASNumAttrs <- tmp$MAS
  # Maximum Edge Value (MEV)
  #statsMEVNumAttrs <- statsMEV_NumAttrs(ds,tgt)
  statsMEVNumAttrs <- tmp$MEV
  # Minimum Cell Number (MCN)
  #statsMCNNumAttrs <- statsMCN_NumAttrs(ds,tgt)
  statsMCNNumAttrs <- tmp$MCN
  # Total Information Coefficient (TIC)
  #statsTICNumAttrs <- statsTIC_NumAttrs(ds,tgt)
  statsTICNumAttrs <- tmp$TIC
  # Entropy
  statsEntropy <- statsEnt(ds,tgt)

  # Mutual Information of Attributes
  statsMUI <- statsMuI(ds,tgt)

  # Measures of Overlapping
  mOverlap <- measOverlap(ds,tgt)

  # Percentual Difference between Classes
  pdc <- ClassPD(ds,form,numCores=numCores)

  landmarkers.tree1 <- landmarker.tree(ds, form, maxdepth = 1)
  landmarkers.tree2 <- landmarker.tree(ds, form, maxdepth = 2)
  landmarkers.tree3 <- landmarker.tree(ds, form, maxdepth = 3)
  landmarkers.nb <- landmarker.nb(ds, form)

  df <- data.frame(NumCases=ncases,
                   imbRatio=imbRatio,
                   NumAttributes=nattributes,
                   NumericVars=numvars,
                   NominalVars=nomvars,
                   ratioCasesAttributes=ratioCasesAttr,
                   attrWithOutliers=attributesWithOutliers,
                   StatsIQR=statsIQR,
                   StatsCoefVar=statsCoefVar,
                   StatsCorBetweenNumVars=statsCorNumAttrs,
                   StatsGKurBetweenNumVars=statsGKurNumAttrs,
                   StatsPKurBetweenNumVars=statsPKurNumAttrs,
                   StatsSkewBetweenNumVars=statsSkewNumAttrs,
                   StatsMICBetweenNumVars=statsMICNumAttrs,
                   StatsMASBetweenNumVars=statsMASNumAttrs,
                   StatsMEVBetweenNumVars=statsMEVNumAttrs,
                   StatsMCNBetweenNumVars=statsMCNNumAttrs,
                   StatsTICBetweenNumVars=statsTICNumAttrs,
                   StatsEntropy=statsEntropy,
                   StatsMutualInfoBetweenNumVars=statsMUI,
                   OverlapMeasures=mOverlap,
                   ClassDiff=pdc,
                   lmkTree1=landmarkers.tree1,
                   lmkTree2=landmarkers.tree2,
                   lmkTree3=landmarkers.tree3,
                   lmkNB=landmarkers.nb)

  df

}