R/classification_metric.R

In aif360: Help Detect and Mitigate Bias in Machine Learning Models

#' Classification Metric
#' @description
#' Class for computing metrics based on two BinaryLabelDatasets. The first dataset is the original one and the second is the output of the classification transformer (or similar)
#' @param dataset (BinaryLabelDataset) Dataset containing ground-truth labels
#' @param classified_dataset (BinaryLabelDataset) Dataset containing predictions
#' @param privileged_groups Privileged groups. List containing privileged protected attribute name and value of the privileged protected attribute.
#' @param unprivileged_groups Unprivileged groups. List containing unprivileged protected attribute name and value of the unprivileged protected attribute.
#' @usage
#' classification_metric(dataset, classified_dataset, unprivileged_groups, privileged_groups)
#' @examples
#' \donttest{
#' load_aif360_lib()
#' # Input dataset
#' data <- data.frame("feat" = c(0,0,1,1,1,1,0,1,1,0), "label" = c(1,0,0,1,0,0,1,0,1,1))
#' # Create aif compatible input dataset
#' act <- aif360::aif_dataset(data_path = data,  favor_label=0, unfavor_label=1,
#'                             unprivileged_protected_attribute=0,
#'                             privileged_protected_attribute=1,
#'                             target_column="label", protected_attribute="feat")
#' # Classified dataset
#' pred_data <- data.frame("feat" = c(0,0,1,1,1,1,0,1,1,0), "label" = c(1,0,1,1,1,0,1,0,0,1))
#' # Create aif compatible classified dataset
#' pred <- aif360::aif_dataset(data_path = pred_data,  favor_label=0, unfavor_label=1,
#'                              unprivileged_protected_attribute=0,
#'                              privileged_protected_attribute=1,
#'                              target_column="label", protected_attribute="feat")
#' # Create an instance of classification metric
#' cm <- classification_metric(act, pred, list('feat', 1), list('feat', 0))
#' # Access metric functions
#' cm$accuracy()
#' }
#' @seealso
#' \href{https://aif360.readthedocs.io/en/latest/modules/metrics.html#classification-metric}{Explore available classification metrics explanations here}
#'
#' Available metrics:
#' \itemize{
#'   \item accuracy
#'   \item average_abs_odds_difference
#'   \item average_odds_difference
#'   \item between_all_groups_coefficient_of_variation
#'   \item between_all_groups_generalized_entropy_index
#'   \item between_all_groups_theil_index
#'   \item between_group_coefficient_of_variation
#'   \item between_group_generalized_entropy_index
#'   \item between_group_theil_index
#'   \item binary_confusion_matrix
#'   \item coefficient_of_variation
#'   \item disparate_impact
#'   \item equal_opportunity_difference
#'   \item error_rate
#'   \item error_rate_difference
#'   \item error_rate_ratio
#'   \item false_discovery_rate
#'   \item false_discovery_rate_difference
#'   \item false_discovery_rate_ratio
#'   \item false_negative_rate
#'   \item false_negative_rate_difference
#'   \item false_negative_rate_ratio
#'   \item false_omission_rate
#'   \item false_omission_rate_difference
#'   \item false_omission_rate_ratio
#'   \item false_positive_rate
#'   \item false_positive_rate_difference
#'   \item false_positive_rate_ratio
#'   \item generalized_binary_confusion_matrix
#'   \item generalized_entropy_index
#'   \item generalized_false_negative_rate
#'   \item generalized_false_positive_rate
#'   \item generalized_true_negative_rate
#'   \item generalized_true_positive_rate
#'   \item negative_predictive_value
#'   \item num_false_negatives
#'   \item num_false_positives
#'   \item num_generalized_false_negatives
#'   \item num_generalized_false_positives
#'   \item num_generalized_true_negatives
#'   \item num_generalized_true_positives
#'   \item num_pred_negatives
#'   \item num_pred_positives
#'   \item num_true_negatives
#'   \item num_true_positives
#'   \item performance_measures
#'   \item positive_predictive_value
#'   \item power
#'   \item precision
#'   \item recall
#'   \item selection_rate
#'   \item sensitivity
#'   \item specificity
#'   \item statistical_parity_difference
#'   \item theil_index
#'   \item true_negative_rate
#'   \item true_positive_rate
#'   \item true_positive_rate_difference
#'
#' }
#' @export
#' @importFrom reticulate py_suppress_warnings
#'
classification_metric <- function(dataset,
                                  classified_dataset,
                                  unprivileged_groups,
                                  privileged_groups){

  u_dict <- dict_fn(unprivileged_groups)

  p_dict <- dict_fn(privileged_groups)


  return(metrics$ClassificationMetric(dataset,
                                      classified_dataset,
                                      unprivileged_groups = u_dict,
                                      privileged_groups = p_dict))
}