R/correl_feature_elimination.R

In elpidiofilho/mdsFuncs: Auxiliar functions to pre-process raster data and modelling with machine learning models using caret package

#' High correlation feature
#' @description select variables with correlation is greater then
#'    a threshold value
#' @param mcor matrix correlation matrix
#' @param testcor numeric values of max correlation to be tested
#' @param graph boolean plot graphic
#' @import ggplot2
#' @return character vector with features names
#' @export
#' @examples
#' # cor_high = correl_feat_elimin(mcor, testcor = c(0.80, 0.90, 0.95, 0.975))
correl_feat_elimin <- function(mcor, testcor = c(0.80, 0.85, 0.90, 0.925, 0.95, 0.975),
                               graph = TRUE) {
  var_ <- correl <- NULL
  df <- data.frame()
  nmax <- length(testcor)
  i <- 1
  flag <- FALSE
  for (i in 1:nmax) {
    vc <- caret::findCorrelation(mcor, cutoff = testcor[i], names = TRUE)
    if (length(vc) == 0) {
      next
    } else {
      if (flag == FALSE) {
        df1 <- data.frame(var_ = vc, correl = testcor[i])
        df <- df1
        flag <- TRUE
      } else {
        df1 <- data.frame(var_ = vc, correl = testcor[i])
        df <- rbind(df, df1)
      }
    }
  }
  print(nrow(df))
  if (!IsEqual(nrow(df), 0)) {
    dfremove <- df |>
      dplyr::group_by(var_) |>
      dplyr::slice(which.max(correl))
    if (graph == TRUE) {
      g <- ggplot(dfremove, aes(x = correl, y = reorder(var_, correl))) +
        geom_col(show.legend = FALSE, fill = "light blue") +
        geom_text(aes(label = correl)) +
        labs(
          x = "Threshold correlation", y = "Eliminated features",
          title = "Features elimination by correlation"
        ) +
        theme_bw()
      print(g)
    }
    return(dfremove)
  } else {
    warning("No features with high correlation")
    return(NULL)
  }
}