In tkonopka/mlensemble: Ensembles of Machine Learning Models

Hooks

Hooks are functions that are executed at predefined points in a pipeline. In the context of a machine learning pipelines, such functions can be used to pre-process data before applying a machine learning model, or to post-process outputs.

As an example, consider that future data may be supplied in a different format with variables denoted as X1 and X2 instead of the names x1 and x2 used for training.

data_uppercase <- data.frame(
  X1 = seq(1, 4) + rnorm(4, 0, 0.2),
  X2 = seq(1, 4) + rnorm(4, 0, 0.2)
)

Direct application models trained on dataset with variables x1 and x2 is bound to give either errors or incomplete outputs. For example,

predict(e_lm_calib, data_uppercase)

To make prediction on such new datasets, the data should be adjusted before performing the prediction. This transformation can be incorporated into an ml_model or ml_ensemble via a hook.

cols_lowercase <- function(x) {
  colnames(x) <- tolower(colnames(x))
  x
}
hook_lowercase <- ml_hook(cols_lowercase, type="pre")
hook_lowercase

The function cols_lowercase takes one argument and returns one object. The hook hook_lowercase is defined specifying that the function should be used as pre-processing step.

A list of hooks can be attached to a model in the ml_model constructor.

e_models_with_hooks <-
  ml_model(lm(y ~ x1, data=data_train), name="lm_x1", hooks=list(hook_lowercase)) +
  ml_model(lm(y ~ x2, data=data_train), name="lm_x2", hooks=list(hook_lowercase))

This construction is reasonable if models require different sets of pre- and post-processing steps. When a step can be applied just once, then a hook can be associated with the ensemble instead.

e_with_hooks <- ml_ensemble(hooks=list(hook_lowercase)) +
        ml_model(lm(y ~ x1, data=data_train), name="lm_x1") +
        ml_model(lm(y ~ x2, data=data_train), name="lm_x2")
e_with_hooks_calib <- calibrate(e_with_hooks, data_calib, label=data_calib$y)
predict(e_with_hooks_calib, data_uppercase)

Thanks to the pre-processing transformation, the ensemble generates predictions even though the new data is not in the same format as the training data.

For further details on how to define hooks, see help(ml_hook). For details on how to attach hooks, see help(ml_model) and help(ml_ensemble).

tkonopka/mlensemble documentation built on March 19, 2022, 7:28 a.m.