R/get.importance.R

In software-analytics/Rnalytica: An R package of JIRA defect datasets and tool suites for explainable software analytics

#' Get variable importance for Logistic Regression and Random Forest
#'
#' This function makes life simple by automatically interpreting logistic regression and random forest models.
#' The interpretation techniques for logistic regression are ANOVA Type-I and variants of ANOVA Type-II (i.e., Wald, Chisq, and F).
#' The interpretation techniques for random forest are Gini and Permutation (i.e., Scaled and Raw) importance.
#'
#' @param model a logistic regression or random forest model
#' @param normalized a logical to indicate whether to normalize importance scores

#' @importFrom stats anova
#' @keywords interpretation
#' @export
get.importance <- function(model, normalized = F) {
  model.technique <- class(model)[1]
  if (model.technique == 'glm') {
    # Logistic regression

    indep <- names(model$coefficients)[-1]
    # Type-I
    anova1 <- anova(model)$Deviance[-1]
    names(anova1) <- indep
    anova1[is.na(anova1)] <- 0

    # Type-II (i.e., Wald, Chisq, and F)
    anova2.wald <-
      Anova(
        model,
        type = 2,
        test.statistic = 'Wald',
        singular.ok = T
      )[, 2]
    names(anova2.wald) <- indep
    anova2.wald[is.na(anova2.wald)] <- 0

    anova2.chisq <- Anova(
      model,
      type = 2,
      test.statistic = 'LR',
      singular.ok = T
    )[, 1]
    names(anova2.chisq) <- indep # label
    anova2.chisq[is.na(anova2.chisq)] <- 0

    anova2.f <- Anova(
      model,
      type = 2,
      test.statistic = 'F',
      singular.ok = T
    )[1:length(indep), 3]
    names(anova2.f) <- indep
    anova2.f[is.na(anova2.f)] <- 0 # remove NA

    # Normalize
    if (normalized) {
      anova1 <-
        anova1 / sum(anova1) * 100
      anova2.wald <-
        anova2.wald / sum(anova2.wald) * 100
      anova2.chisq <-
        anova2.chisq / sum(anova2.chisq) * 100
      anova2.f <-
        anova2.f / sum(anova2.f) * 100
    }

    return(
      list(
        ANOVA.TypeI = anova1,
        ANOVA.TypeII.Wald = anova2.wald,
        ANOVA.TypeII.Chisq = anova2.chisq,
        ANOVA.TypeII.F = anova2.f
      )
    )
  } else if (model.technique == 'randomForest') {
    # Random forest
    importance <- ranger::importance(model)

    # Normalize
    if (normalized) {
      importance <- importance/sum(importance)*100
    }
    return(list(
      importance = importance
    ))
  } else {
    stop('Input model must be Logistic Regression or Random Forest')
  }
}