Fit a glmnet model and calculate performance metrics using a single rsplit object
Source:R/modeling_helpers.R
tof_fit_split.RdThis function trains a glmnet model on the training set of an rsplit object, then calculates performance metrics of that model on the validation/holdout set at all combinations of the mixture and penalty hyperparameters provided in a hyperparameter grid.
Usage
tof_fit_split(
split_data,
prepped_recipe,
hyperparameter_grid,
model_type,
outcome_colnames
)Arguments
- split_data
An `rsplit` object from the
rsamplepackage. Alternatively, an unsplit tbl_df can be provided, though this is not recommended.- prepped_recipe
A trained
recipe- hyperparameter_grid
A tibble containing the hyperparameter values to tune. Can be created using
tof_create_grid- model_type
A string representing the type of glmnet model being fit.
- outcome_colnames
Quoted column names indicating which columns in the data being fit represent the outcome variables (with all others assumed to be predictors).
Value
A tibble with the same number of rows as the input hyperparameter grid. Each row represents a combination of mixture and penalty, and each column contains a performance metric for the fitted glmnet model on `split_data`'s holdout set. The specific performance metrics depend on the type of model being fit:
- "linear"
mean-squared error (`mse`) and mean absolute error (`mae`)
- "two-class"
binomial deviance (`binomial_deviance`); misclassification error rate `misclassification_error`; the area under the receiver-operating curve (`roc_auc`); and `mse` and `mse` as above
- "multiclass"
multinomial deviance (`multinomial_deviance`); misclassification error rate `misclassification_error`; the area under the receiver-operating curve (`roc_auc`) computed using the Hand-Till method in
roc_auc; and `mse` and `mse` as above- "survival"
the negative log2-transformed partial likelihood (`neg_log_partial_likelihood`) and Harrel's concordance index (often simply called "C"; `concordance_index`)