Skip to contents

Tune a deweather model

Source: R/tune_dw_model.R
tune_dw_model.Rd

This function performs hyperparameter tuning for a gradient boosting model used in deweathering air pollution data. It uses cross-validation to find optimal hyperparameters and returns the best performing model along with performance metrics and visualizations. Parallel processing (e.g., through the mirai package) is recommended to speed up tuning - see https://tune.tidymodels.org/articles/extras/optimizations.html#parallel-processing.

Usage

tune_dw_model(
  data,
  pollutant,
  vars = c("trend", "ws", "wd", "hour", "weekday", "air_temp"),
  tree_depth = 5,
  trees = 200L,
  learn_rate = 0.1,
  mtry = NULL,
  min_n = 10L,
  loss_reduction = 0,
  sample_size = 1L,
  stop_iter = 190L,
  engine = c("xgboost", "lightgbm"),
  split_prop = 3/4,
  grid_levels = 5,
  v_partitions = 10
)

Arguments

data

An input data.frame containing one pollutant column (defined using pollutant) and a collection of feature columns (defined using vars).

pollutant

The name of the column (likely a pollutant) in data to predict.

vars

The name of the columns in data to use as model features - i.e., to predict the values in the pollutant column. Any character columns will be coerced to factors. "hour", "weekday", "trend", "yday", "week", and "month" are special terms and will be passed to append_dw_vars() if not present in names(data).

tree_depth, trees, learn_rate, mtry, min_n, loss_reduction, sample_size, stop_iter

If length 1, these parameters will be fixed. If length 2, the parameter will be tuned within the range defined between the first and last value. For example, if tree_depth = c(1, 5) and grid_levels = 3, tree depths of 1, 3, and 5 will be tested.

engine

A single character string specifying what computational engine to use for fitting.

split_prop

The proportion of data to be retained for modeling/analysis. Passed to the prop argument of rsample::initial_split().

grid_levels

An integer for the number of values of each parameter to use to make the regular grid. Passed to the levels argument of dials::grid_regular().

v_partitions

The number of partitions of the data set to use for v-fold cross-validation. Passed to the v argument of rsample::vfold_cv().

Details

The function performs the following steps:

  • Removes rows with missing values in the pollutant or predictor variables

  • Splits data into training and testing sets

  • Creates a tuning grid for any parameters specified as ranges

  • Performs grid search with cross-validation to find optimal hyperparameters

  • Fits a final model using the best hyperparameters

  • Generates predictions and performance metrics

    At least one hyperparameter must be specified as a range (vector of length 2) for tuning to occur. Single values are treated as fixed parameters.

Author

Jack Davison