bst_control: Control Parameters for Boosting

In bst: Gradient Boosting

Control Parameters for Boosting

Description

Specification of the number of boosting iterations, step size and other parameters for boosting algorithms.

Usage

bst_control(mstop = 50, nu = 0.1, twinboost = FALSE, twintype=1, threshold=c("standard", 
"adaptive"), f.init = NULL, coefir = NULL, xselect.init = NULL, center = FALSE, 
trace = FALSE, numsample = 50, df = 4, s = NULL, sh = NULL, q = NULL, qh = NULL, 
fk = NULL, start=FALSE, iter = 10, intercept = FALSE, trun=FALSE)

Arguments

mstop	an integer giving the number of boosting iterations.
nu	a small number (between 0 and 1) defining the step size or shrinkage parameter.
twinboost	a logical value: TRUE for twin boosting.
twintype	for twinboost=TRUE only. For learner="ls", if twintype=1, twin boosting with weights from magnitude of coefficients in the first round of boosting. If twintype=2, weights are correlations between predicted values in the first round of boosting and current predicted values. For learners not componentwise least squares, twintype=2.
threshold	if threshold="adaptive", the estimated function ctrl$fk is updated in every boosting step. Otherwise, no update for ctrl$fk in boosting steps. Only used in robust nonconvex loss function.
f.init	the estimate from the first round of twin boosting. Only useful when twinboost=TRUE and learner="sm" or "tree".
coefir	the estimated coefficients from the first round of twin boosting. Only useful when twinboost=TRUE and learner="ls".
xselect.init	the variable selected from the first round of twin boosting. Only useful when twinboost=TRUE.
center	a logical value: TRUE to center covariates with mean.
trace	a logical value for printout of more details of information during the fitting process.
numsample	number of random sample variable selected in the first round of twin boosting. This is potentially useful in the future implementation.
df	degree of freedom used in smoothing splines.
s,q	nonconvex loss tuning parameter s or frequency q of outliers for robust regression and classification. If s is missing but q is available, s may be computed as the 1-q quantile of robust loss values using conventional software.
sh, qh	threshold value or frequency qh of outliers for Huber regression family="huber" or family="rhuberDC". For family="huber", if sh is not provided, sh is then updated adaptively with the median of y-yhat where yhat is the estimated y in the last boosting iteration. For family="rhuberDC", if sh is missing but qh is available, sh may be computed as the 1-qh quantile of robust loss values using conventional software.
fk	predicted values at an iteration in the MM algorithm
start	a logical value, if start=TRUE and fk is a vector of values, then bst iterations begin with fk. Otherwise, bst iterations begin with the default values. This can be useful, for instance, in rbst for the MM boosting algorithm.
iter	number of iteration in the MM algorithm
intercept	logical value, if TRUE, estimation of intercept with linear predictor model
trun	logical value, if TRUE, predicted value in each boosting iteration is truncated at -1, 1, for family="closs" in bst and rfamily="closs" in rbst

Details

Objects to specify parameters of the boosting algorithms implemented in bst, via the ctrl argument. The s value is for robust nonconvex loss where smaller s value is more robust to outliers with family="closs", "tbinom", "thinge", "tbinomd", and larger s value more robust with family="clossR", "gloss", "qloss".

For family="closs", if s=2, the loss is similar to the square loss; if s=1, the loss function is an approximation of the hinge loss; for smaller values, the loss function approaches the 0-1 loss function if s<1, the loss function is a nonconvex function of the margin.

The default value of s is -1 if family="thinge", -log(3) if family="tbinom", and 4 if family="binomd". If trun=TRUE, boosting classifiers can produce real values in [-1, 1] indicating their confidence in [-1, 1]-valued classification. cf. R. E. Schapire and Y. Singer. Improved boosting algorithms using confidence-rated predictions. In Proceedings of the Eleventh Annual Conference on Computational Learning Theory, pages 80-91, 1998.

Value

An object of class bst_control, a list. Note fk may be updated for robust boosting.

See Also

bst

bst documentation built on Jan. 7, 2023, 1:23 a.m.