boostComponents: Wrapper to boost general additive models using components

In schalkdaniel/compboost: Efficient Component-Wise Boosting Implementation

Wrapper to boost general additive models using components

Description

This wrapper function automatically initializes the model by adding all numerical features as components. This means, that for each numerical feature a linear effect and non-linear spline base-learner is added. The non-linear part is constructed in way that it cannot model the linear part. Hence, it is just selected if a non-linear base learner is really necessary. Categorical features are dummy encoded and inserted using another linear base-learners without intercept.

The returned object is an object of the Compboost class. This object can be used for further analyses (see ?Compboost for details).

Usage

boostComponents(
  data,
  target,
  optimizer = NULL,
  loss = NULL,
  learning_rate = 0.05,
  iterations = 100,
  trace = -1,
  degree = 3,
  n_knots = 20,
  penalty = 2,
  df = 0,
  differences = 2,
  data_source = InMemoryData,
  oob_fraction = NULL,
  bin_root = 0,
  cache_type = "inverse",
  stop_args = list(),
  df_cat = 1,
  stop_time = "microseconds",
  additional_risk_logs = list()
)

Arguments

data	(data.frame()) A data frame containing the data.
target	(character(1) | ResponseRegr | ResponseBinaryClassif) Character value containing the target variable or response object. Note that the loss must match the data type of the target.
optimizer	(OptimizerCoordinateDescent | OptimizerCoordinateDescentLineSearch | OptimizerAGBM | OptimizerCosineAnnealing) An initialized S4 optimizer object (requires to call Optimizer*.new(..). See the respective help page for further information.
loss	(LossQuadratic | LossBinomial | LossHuber | LossAbsolute | LossQuantile) An initialized S4 loss object (requires to call ⁠Loss*$new(...)⁠). See the respective help page for further information.
learning_rate	(numeric(1)) Learning rate to shrink the parameter in each step.
iterations	(integer(1)) Number of iterations that are trained. If iterations == 0, the untrained object is returned. This can be useful if other base learners (e.g. an interaction via a tensor base learner) are added.
trace	(integer(1)) Integer indicating how often a trace should be printed. Specifying trace = 10, then every 10th iteration is printed. If no trace should be printed set trace = 0. Default is -1 which means that in total 40 iterations are printed.
degree	(integer(1))cr Polynomial degree of the splines.
n_knots	(integer(1)) Number of equidistant "inner knots". The actual number of used knots does also depend on the polynomial degree.
penalty	(numeric(1)) Penalty term for p-splines. If the penalty equals 0, then ordinary b-splines are fitted. The higher the penalty, the higher the smoothness.
df	(numeric(1)) Degrees of freedom of the base learner(s).
differences	(integer(1)) Number of differences that are used for penalization. The higher the difference, the higher the smoothness.
data_source	(⁠Data*⁠) Uninitialized ⁠Data*⁠ object which is used to store the data. At the moment just in memory training is supported.
oob_fraction	(numeric(1)) Fraction of how much data are used to track the out of bag risk.
bin_root	(integer(1)) The binning root to reduce the data to n^{1/\text{binroot}} data points (default bin_root = 1, which means no binning is applied). A value of bin_root = 2 is suggested for the best approximation error (cf. Wood et al. (2017) Generalized additive models for gigadata: modeling the UK black smoke network daily data).
cache_type	(character(1)) String to indicate what method should be used to estimate the parameter in each iteration. Default is cache_type = "cholesky" which computes the Cholesky decomposition, caches it, and reuses the matrix over and over again. The other option is to use cache_type = "inverse" which does the same but caches the inverse.
stop_args	(list(2)) List containing two elements patience and eps_for_break which can be set to use early stopping on the left out data from setting oob_fraction. If ! is.null(stop_args), early stopping is triggered.
df_cat	(numeric(1)) Degrees of freedom of the categorical base-learner.
stop_time	(character(1)) Unit of measured time.
additional_risk_logs	(list(Logger)) Additional logger passed to the Compboost object.

Value

A model of the Compboost class. This model is an R6 object which can be used for retraining, predicting, plotting, and anything described in ?Compboost.

Examples

mod = boostComponents(data = iris, target = "Sepal.Length", df = 4)
mod$getBaselearnerNames()
table(mod$getSelectedBaselearner())
plotPEUni(mod, "Petal.Length")
mod$predict()

schalkdaniel/compboost documentation built on April 15, 2023, 9:03 p.m.