R/calculate_feature_importances.R

In dynverse/dynfeature: Feature Importance for Dynamic Processes

# Internal documentation for developers only:
# Calculate feature importance scores
#
# @param X A data frame containing the features as columns.
# @param Y A data frame of predictor variables, with `nrow(Y) == nrow(X)`.
# @param fi_method A feature importance method. Default: `fi_ranger_rf_lite()`. Check `?fi_methods` for a full list of available feature importance methods.
# @param verbose Whether to print out extra information.
#
# @returns A data frame with three columns, `predictor_id`, `feature_id`, and `importance`. `predictor_id` is a column in `Y`, while `feature_id` is a column in `X`.
#
# @examples
# X <- data.frame(matrix(runif(25*10), ncol = 10))
# Y <- data.frame(matrix(runif(25*2), ncol = 2))
#
# calculate_feature_importances(X, Y)
calculate_feature_importances <- function(
  X,
  Y,
  fi_method = fi_ranger_rf_lite(),
  verbose = FALSE
) {
  # if Y is a vector or a matrix, turn it into a data frame
  if (!is.data.frame(Y)) {
    # convert to regular matrix if sparse
    if (dynutils::is_sparse(Y)) {
      Y <- as.matrix(Y)
    }
    Y <- as_tibble(Y, .name_repair = "minimal")
  }

  # convert expression to regular matrix if sparse
  if (dynutils::is_sparse(X)) {
    X <- as.matrix(X)
  }

  # calculate importance score for each predictor separately
  importances <- map_df(seq_len(ncol(Y)), function(i) {
    if (verbose) cat("Generating forest ", i , "/", ncol(Y), "\n", sep = "")

    y <- Y %>% pull(!!i)

    if (is.character(y) || is.logical(y)) {
      y <- factor(y)
    }

    # return 0 if y are all the same values
    if (length(unique(y)) == 1) {
      importance <- set_names(rep(0, ncol(X)), colnames(X))
    } else {
      importance <- fi_method$fun(X, y, verbose = verbose)
    }

    tibble(
      predictor_id = factor(colnames(Y)[[i]], levels = colnames(Y)),
      feature_id = factor(names(importance), levels = names(importance)),
      importance
    )
  })

  # return importances ordered by value
  importances %>%
    arrange(desc(.data$importance))
}