R/rmw_partial_dependencies.R

In rmweather: Tools to Conduct Meteorological Normalisation and Counterfactual Modelling for Air Quality Data

#' Function to calculate partial dependencies after training with 
#' \strong{rmweather}. 
#' 
#' \code{rmw_plot_partial_dependencies} is rather slow. 
#' 
#' @param model A ranger model object from \code{\link{rmw_train_model}}. 
#' 
#' @param df Input data frame after preparation with 
#' \code{\link{rmw_prepare_data}}.
#' 
#' @param variable Vector of variables to calculate partial dependencies for. 
#' 
#' @param training_only Should only the training set be used for prediction? The
#' default is \code{TRUE}. 
#' 
#' @param n_cores Number of CPU cores to use for the model calculation. The 
#' default is system's total minus one.
#' 
#' @param resolution The number of points that should be predicted for each 
#' independent variable. If left as \code{NULL}, a default sequence will be 
#' generated. See \code{\link{partial}} for details. 
#' 
#' @param verbose Should the function give messages? 
#' 
#' @return Tibble. 
#' 
#' @author Stuart K. Grange
#' 
#' @examples 
#' 
#' \donttest{
#' 
#' # Load packages
#' library(dplyr)
#' # Ranger package needs to be loaded
#' library(ranger)
#' 
#' # Prepare example data
#' data_london_prepared <- data_london %>% 
#'   filter(variable == "no2") %>% 
#'   rmw_prepare_data()
#' 
#' # Calculate partial dependencies for wind speed
#' data_partial <- rmw_partial_dependencies(
#'   model = model_london, 
#'   df = data_london_prepared, 
#'   variable = "ws", 
#'   verbose = TRUE
#' )
#' 
#' # Calculate partial dependencies for all independent variables used in model
#' data_partial <- rmw_partial_dependencies(
#'   model = model_london, 
#'   df = data_london_prepared, 
#'   variable = NA, 
#'   verbose = TRUE
#' )
#' 
#' }
#' 
#' @export
rmw_partial_dependencies <- function(model, df, variable, training_only = TRUE, 
                                     resolution = NULL, n_cores = NA, 
                                     verbose = FALSE) {
  
  # Check, predict is a generic function and needs to be loaded
  if (!"package:ranger" %in% search()) {
    cli::cli_abort(
      "The ranger package is not loaded and is required for this function."
    )
  }
    
  # Check inputs
  df <- rmw_check_data(df, prepared = TRUE)
  stopifnot(inherits(model, "ranger"))
  
  # Default logic for cpu cores
  n_cores <- as.integer(n_cores)
  n_cores <- if_else(is.na(n_cores), n_cores_default(), n_cores)
  
  # Predict all variables if not given
  if (is.na(variable[1])) {
    variable <- model$forest$independent.variable.names
  }
  
  # Message to user
  if (verbose) {
    cli::cli_alert_info(
      "{str_date_formatted()}: Predicting `{length(variable)}` variable{?s}..."
    )
  }
  
  # Calculate the partial dependencies
  df_predict <- variable %>% 
    purrr::map(
      ~rmw_partial_dependencies_worker(
        model = model,
        df = df, 
        variable = .x,
        training_only = training_only,
        resolution = resolution,
        n_cores = n_cores
      )
    ) %>% 
    purrr::list_rbind() %>% 
    as_tibble()
  
  return(df_predict)
  
}


rmw_partial_dependencies_worker <- function(model, df, variable, training_only, 
                                            resolution, n_cores) {
  
  # Filter only to training set
  if (training_only) {
    df <- filter(df, set == "training")
  }
  
  # Predict
  df_predict <- pdp::partial(
    model,
    pred.var = variable,
    train = df,
    grid.resolution = resolution,
    num.threads = n_cores
  )
  
  # Alter names and add variable
  df_predict <- df_predict %>% 
    purrr::set_names(c("value", "partial_dependency")) %>% 
    as_tibble() %>% 
    mutate(variable = !!variable) %>% 
    relocate(variable, 
             value, 
             partial_dependency)
  
  # Catch factors, usually weekday
  if ("factor" %in% class(df_predict$value)) {
    df_predict$value <- as.integer(as.character(df_predict$value))
  }
  
  return(df_predict)
  
}