soften: Create a 'soft tree' structure with softening parameters set...

In SplitSoftening: Softening Splits in Decision Trees

Create a ‘soft tree’ structure with softening parameters set using one of the named method.

Description

This is a convenience method implemented over softsplits and the softening functions from this package.

Usage

soften(fit, ds, method, control = NULL)

Arguments

fit	A classification tree - either an object tree which represents a classification tree, or already a ‘soft tree’ like created by the softsplits function.
ds	A data frame used as training data for softening
method	A name of softening method. One of: "DR0", "DR1", "DR2", ..., "ESD", "C4.5", "optim_d", "optim_d^2", "optim_d^4", "optim_auc" The 'method = "DRx"' for some number x: The softening parameters are set according to ‘data ranges’ appropriate to tree nodes. The parameters are configured such that in each node the distance of the boundary of the softened area from split value is 2^{-x}r, where r is the distance from the split value to the furthest data point in the tree node projected to the direction from the split value to the boundary. The 'method = "ESD"' sets boundaries of the softening using error standard deviation. This is how C4.5 method sets "probabilistic splits"; for that reason value "C4.5" is an alias for "ESD". The 'method = "optim_d^q"' for some number q: The softening parameters are set by optimization process which minimizes \code{mean}((1.0-p)^q) where p is for each data point in ds the predicted probability of the correct label. If 'method = "optim_auc"': The classification tree fit must perform prediction to two classes. The value of the ‘area under ROC curve’ computed on the data set ds is maximized by optimization.
control	List of additional configuration paramaters. Possible members in the list are: verbosity, implementation, iteration.count, sft.ini, which correspond to the paramaters of softening.optimized.

Value

The ‘soft tree’ structure representing the same tree structure as given in the parameter fit, but with softening parameters set using the given method.

See Also

predictSoftsplits.

SplitSoftening documentation built on Oct. 28, 2024, 1:06 a.m.