Skip to contents

Monotonic constrained priors for boosted regressions

Source: R/prior_gdb.R
GDBPrior.Rd

Monotonic constrains for gradient descent boosting models do not work in the same way as other priors where a specific coefficient or magnitude of importance is specified. Rather monotonic constraints enforce a specific directionality of regression coefficients so that for instance a coefficient has to be positive or negative.

Important: Specifying a monotonic constrain for the engine_gdb does not guarantee that the variable is retained in the model as it can still be regularized out.

Usage

GDBPrior(variable, hyper = "increasing", ...)

# S4 method for class 'character'
GDBPrior(variable, hyper = "increasing", ...)

Arguments

variable

A character matched against existing predictors variables.

hyper

A character object describing the type of constrain. Available options are 'increasing', 'decreasing', 'convex', 'concave', 'positive', 'negative' or 'none'.

...

Variables passed on to prior object.

Value

A Prior object.

Note

Similar priors can also be defined for the engine_xgboost via XGBPrior().

References

  • Hofner, B., Müller, J., & Hothorn, T. (2011). Monotonicity‐constrained species distribution models. Ecology, 92(10), 1895-1901.

See also

Prior, XGBPrior

Other prior: BARTPrior(), BARTPriors(), BREGPrior(), BREGPriors(), GDBPriors(), GLMNETPrior(), GLMNETPriors(), INLAPrior(), INLAPriors(), STANPrior(), STANPriors(), XGBInteractionPrior(), XGBInteractionPriors(), XGBPrior(), XGBPriors(), add_priors(), get_priors(), priors(), rm_priors()

Examples

prior <- GDBPrior("forest", hyper = "increasing")
prior$get("value")
#> [1] "increasing"