internal_workflow: A learning and prediction workflow with internal validation

In mrfoliveira/STResampling-JDSA2020: Biased Resampling Strategies for Imbalanced Spatio-Temporal Forecasting

Description

A learning and prediction workflow that may deal with NAs and use internal validation to parametrize a re-sampling technique to balance an imbalanced regression problem.

Usage

internal_workflow(
  train,
  test,
  form,
  model,
  time,
  site_id,
  resample.grid,
  resample.pars = NULL,
  internal.est = NULL,
  internal.est.pars = NULL,
  internal.evaluator = "int_util_evaluate",
  internal.eval.pars = NULL,
  metrics = c("F1.u", "rmse_phi"),
  metrics.max = c(TRUE, FALSE),
  stat = "MED",
  handleNAs = "centralImputNAs",
  min_train = 2,
  nORp = 0.2,
  .int_parallel = FALSE,
  .intRes = TRUE,
  .full_intRes = FALSE,
  ...
)

Arguments

train	a data frame for training
test	a data frame for testing
form	a formula describing the model to learn
model	the name of the algorithm to use
time	the name of the column in train and test containing time-stamps
site_id	the name of the column in train and test containing location IDs
resample.grid	a data.frame with columns indicating resample.pars to test using internal.est. Any NA value in resample.grid will have the argument set to NULL.
resample.pars	parameters to be passed to re-sample function. Default is NULL.
internal.est	character string identifying the internal estimator function to use
internal.est.pars	named list of internal estimator parameters (e.g., tr.perc or nfolds)
internal.evaluator	character string indicating internal evaluation function
internal.eval.pars	named list of parameters to feed to internal evaluation function
metrics	vector of names of two metrics to be used to determine the best parametrization (the second metric is only used in case of ties)
metrics.max	vector of Booleans indicating whether each metric in parameter metrics should be maximized (TRUE) or minimized (FALsE) for best results
stat	parameter indicating summary statistic that should be used to determine the best internal evaluation metric: "MED" (for median) or "MEAN" (for mean)
handleNAs	string indicating how to deal with NAs. If "centralImputNAs", training observations with at least 80% of non-NA columns, will have their NAs substituted by the mean value and testing observatiosn will have their NAs filled in with mean value regardless. Default is NULL.
min_train	a minimum number of observations that must be left to train a model. If there are not enough observations, predictions will be NA. Default is 2.
nORp	a maximum number or fraction of columns/rows with missing values above which a row/column will be removed from train before learning the model. Only works if handleNAs was set to centralImputNAs. Default is 0.2.
.int_parallel	a Boolean indicating whether rows in the grid search should be tested in parallel
.intRes	a Boolean indicating whether the evalRes object outputed by internal validation should be returned. Defaults to TRUE
.full_intRes	a Boolean indicating whether the full results object for internal validation should be returned as well. Defaults to FALSE
...	other parameters to feed to model

Value

a data frame containing time-stamps, location IDs, true values and predicted values

mrfoliveira/STResampling-JDSA2020 documentation built on June 28, 2021, 7:01 p.m.