Setup NNLS Hyperparameters — setup

Setup hyperparameters for non-negative least squares.

Usage

setup_NNLS(normalize = TRUE, ifw = FALSE)

Arguments

normalize: (Tunable) Logical: If TRUE, scale the coefficients to sum to 1.
ifw: (Tunable) Logical: If TRUE, use Inverse Frequency Weighting in classification.

Value

NNLSHyperparameters object.

Details

NNLS fits y ~ Xb subject to b >= 0, with no intercept. It exists as the default meta learner of the stacked meta learners (setup_SuperLearner, setup_ModalityStacking), where the predictors are the base learners' cross-validated predictions and a non-negative, sum-to-one coefficient vector is the ensemble weighting. It is a poor general-purpose learner: with no intercept and a sign constraint it can only fit outcomes that the predictors already span.

With normalize = TRUE the coefficients are scaled to sum to 1, so the fit is a convex combination of the predictors. For classification the outcome is coded 0/1 on the second factor level and the fitted values are read as probabilities of that level; unnormalized coefficients do not guarantee a value in [0, 1], so predictions are clamped.

Get more information from nnls::nnls.

Author

EDG

Examples

nnls_hyperparams <- setup_NNLS(normalize = FALSE)
nnls_hyperparams
#> <NNLSHyperparameters>
#>         hyperparameters: 
#>                          normalize: <lgc> FALSE
#>                                ifw: <lgc> FALSE
#> tunable_hyperparameters: <chr> normalize, ifw
#>   fixed_hyperparameters: <chr> 
#>                   tuned: <int> -1
#>               resampled: <int> 0
#>               n_workers: <int> 1
#> 
#>   No search values defined for tunable hyperparameters.