trainModel: Model training and performance using cross-validation

In environmentalinformatics-marburg/gpm: Geospatial predictive modeling using parameterized and unparameterized models.

Description

Train a model and estimate the model performance using multiple resamplings each devided into training and independent subsets. Training subsets are further divided into k-fold cross-validation samples for model tuing. Testing sampels are used for the independent validation of the final model. This procedure is repeated for each resampling provided.

Usage

## S4 method for signature 'GPM'
trainModel(x, n_var = NULL, mthd = "rf", mode = c("rfe",
  "ffs"), seed_nbr = 11, cv_nbr = NULL, var_selection = c("sd", "indv"),
  metric = NULL, tune_length = NULL, response_nbr = NULL,
  resample_nbr = NULL, filepath_tmp = NULL, rerank = FALSE, ...)

## S4 method for signature 'data.frame'
trainModel(x, response, predictor, selector, meta,
  resamples, n_var = NULL, mthd = "rf", mode = c("rfe", "ffs"),
  seed_nbr = 11, cv_nbr = 2, var_selection = c("sd", "indv"),
  metric = NULL, tune_length = NULL, response_nbr = NULL,
  resample_nbr = NULL, filepath_tmp = NULL, rerank = FALSE, ...)

Arguments

x	An object of class gpm or data.frame
n_var	Vector holding the number of variables used for the recursive feature elimination iterations; must not be continous (e.g. c(1:10, 20, 30))
mthd	Core method used for the model (e.g. "rf" for random forest)
mode	Variable selection mode, either recursive feature elimination ("rfe") or forward feature selection ("ffs)
seed_nbr	Specific seed to be to ensure reproducability
cv_nbr	Specific cross validation folds to be used for model tuning within each forward or backward feature selection/elimination step
var_selection	Select final number of variables based on a standard deviation statistic ("sd", more conservative) or by the actual best number ("indv")
metric	The metric to be used to compute the model performance.
tune_length	Tune length to be used in recursive feature elimination (if NULL, the fixed default grid taken from the GPM LUT will be used).
response_nbr	Response ID to be computed; only relevant if more than one response variable is present and a model should not be built for each of them
resample_nbr	Resample ID to be computed; only relevant if the model training should not run over all resamples
filepath_tmp	If set, intermediate model results during the variable selection are writen to disc; if the procedure stops for some reason, the already computed results can be read in again which saves computation time (e.g. after an accidential shutdown etc.)
...	Additional arguments passed to trainModelffs.
response	The column name(s) of the response variable(s)
predictor	The column ID of the predictor, i.e. independent variable(s) in the dataset
selector	Selector id
meta	Meta information of the gpm object
resamples	The list of the resamples containing the individual row numbers (resulting from function resamplingsByVariable)

Details

The backfard feature selection is based on the implementation of the caret::rfe function. The forward feature selection is implemented from scratch. The latter stops if the error statistics get worse after a first optimum is reached. For model training, the respective caret functions are used, too.

Value

NONE

A layer within the gpm object with the model training information for each response variable and all resamplings.

Trained model for each response variable and all resamplings.

References

The function uses functions from: Max Kuhn. Contributions from Jed Wing, Steve Weston, Andre Williams, Chris Keefer, Allan Engelhardt, Tony Cooper, Zachary Mayer, Brenton Kenkel, the R Core Team, Michael Benesty, Reynald Lescarbeau, Andrew Ziem, Luca Scrucca, Yuan Tang and Can Candan. (2016). caret: Classification and Regression Training. https://CRAN.R-project.org/package=caret

See Also

NONE

Examples

## Not run: 
#Not run

## End(Not run)

environmentalinformatics-marburg/gpm documentation built on July 11, 2020, 11:12 a.m.