emil (version 2.2.10)

modeling_procedure: Setup a modeling procedure

Description

A modeling procedure is an object containing all information necessary to carry out and evaluate the performance of a predictive modeling task with fit, tune, or evaluate. To use an out-of-the box algorithm with default values, only the method argument needs to be set. See emil for a list of available methods. To deviate from the defaults, e.g. by tuning parameters or using a custom function for model fitting, set the appropriate parameters as described below. For a guide on how to implement a custom method see the documentaion page extension.

Usage

modeling_procedure(method, parameter = list(), error_fun = NULL, fit_fun,
  predict_fun, importance_fun)

Arguments

method

The name of the modeling method. Only needed to identify plug-in functions, i.e. if you supply them yourself there is no need to set method.

parameter

A list of model parameters. These will be fed to the fitting function after the dataset (x and y parameters). To tune a parameter, supply the candidate values in a vector or list.

When tuning more than one parameter, all combinations of parameter values will be tested, if the elements of parameter are named. To manually specify which parameter value combinations to try, leave the the elements unnamed (see example 3 and 4).

Parameters that should have vectors or lists as values, e.g. trControl when using fit_caret to train pkgcaret models, must be wrapped in an additional list. That is, to set a parameter value to a list, but not tune it, make it a list of length 1 containing the list to be used (see example 6).

error_fun

Performance measure used to evaluate procedures and to tune parameters. See error_fun for details.

fit_fun

The function to be used for model fitting.

predict_fun

The function to be used for model prediction.

importance_fun

The function to be used for calculating or extracting feature importances. See get_importance for details.

Value

An object of class modeling_procedure.

See Also

emil, evaluate, fit, tune, predict, get_importance

Examples

# NOT RUN {
# 1: Fit linear discriminants without tuning any parameter,
# since it has none
modeling_procedure("lda")

# 2: Tune random forest's `mtry` parameter, with 3 possible values
modeling_procedure("randomForest", list(mtry = list(100, 250, 1000)))

# 3: Tune random forest's `mtry` and `maxnodes` parameters simultaneously,
# with 3 values each, testing all 9 possible combinations
modeling_procedure("randomForest", list(mtry = list(100, 250, 1000),
                                        maxnodes = list(5, 10, 25)))

# 4: Tune random forest's `mtry` and `maxnodes` parameters simultaneously,
# but only test 3 manually specified combinations of the two
modeling_procedure("randomForest", list(list(mtry = 100, maxnodes = 5),
                                   list(mtry = 250, maxnodes = 10),
                                   list(mtry = 1000, maxnodes = 25)))

# 5: Tune elastic net's `alpha` and `lambda` parameters. Since elastic net's
# fitting function can tune `lambda` internally in a more efficient way
# than the general framework is able to do, only tune `alpha` and pass all
# `lambda` values as a single argument.
modeling_procedure("glmnet", list(alpha = seq(0, 1, length.out=6),
                                  lambda = list(seq(0, 5, length.out=30))))

# 6: Train elastic nets using the caret package's model fitting framework
if(requireNamespace("caret", quitely = TRUE)){
    modeling_procedure("caret", list(method = "glmnet",
        trControl = list(trainControl(verboseIter = TRUE, classProbs = TRUE))))
}

# }