rf.boot: Bootstrap analysis for null hypothesis testing of random...

In raytut/randomForestLOO: Random Forests with Leave-One-Out

Bootstrap analysis for null hypothesis testing of random forests with Leave-One-Out procedures.

Description

Function to run a bootstrap analysis with leave-one-out analysis. Returns a dataframe with all bootstrap errors for the estimations, and the average errors per subject.

Usage

rf.boot(data, sub_id, xvar, yvar, method, ntree, iterations, ncore)

Arguments

data	Dataframe to be used by model
sub_id	String indicating what variable to use as subject ID
xvar	Vector of strings to be used for model training
yvar	A response string. Must be of type factor.
method	String, can be "LOSO" or "LOBO". Default="LOBO".
ntree	Integer indicating number of trees to simulate. Default=1000.
iterations	Number of times to run bootstrap. Default=5000.
ncore	Number of cores assigned for parallel processing. Default=1 (Not recommended). Will always use n-1 to ensure at least one core remains available and free for smooth operaitons.

Value

Results of the bootstrap analysis as a list:

model	Information on the model that was run
subject	Results for each subject, including the predictions made, a confusion matrix, and the mean error
errors	Mean error rate returned per subject
mean_errors	Mean error rate across subjects

Note

Currenlty implemented only for factors. Random forests are not optimal for regressions, limiting the utility of such applicaitons. Note that sometimes the total number of iterations may be slightly different. This is because failed bootstraps are discarded from the final dataframe.

Author(s)

Rayyan Tutunji | rayyan.tutunji[at]donders.ru.nl

References

Using wearable biosensors and ecological momentary assessments for the detection of prolonged stress in real life Rayyan Tutunji, Nikos Kogias, Bob Kapteijns, Martin Krentz, Florian Krause, Eliana Vassena, Erno Hermans bioRxiv 2021.06.29.450360; doi: https://doi.org/10.1101/2021.06.29.450360

raytut/randomForestLOO documentation built on May 30, 2022, 8:47 p.m.