gini_index: compute Gini impurity for binary values only

In markusloecher/rfVarImpOOB: Unbiased Variable Importance for Random Forests

Description

simple function to compute simple or penalized Gini impurity

The "penalty" compares the class probabilities pHat with a reference estimate pEst

which would typically serve as a prediction (e.g. in a tree node).

Usage

gini_index(pHat, pEst = NULL, k = 2, kind = 1, w = 2, 


    correctBias = FALSE, nNode = NULL, verbose = 0)

Arguments

pHat	probabilities from the current data,
pEst	estimated class probabilities (typically from an earlier inbag estimation). Only pass if you intend to compute the "validation-penalized Gini"
k	exponent of penalty term: abs(pHat-pEst)^k
kind	kind of penalty
w	weights, default is 2 if you pass just a single probability instead of the vector (p,1-p),
correctBias	multiply by n/(n-1) for sample variance correction!
nNode	number of observations in node; only used for one special Gini impurity!
verbose	level of verbosity

Value

simple or penalized Gini impurity

Author(s)

Markus Loecher <[e-mail]>

Examples

#Test binary case:





gini_index(0.5,0.5,kind=1)


gini_index(0.9,0.1,kind=1)


gini_index(0.1,0.9,kind=1)





gini_index(0.5,0.5,kind=2)


gini_index(0.9,0.1,kind=2)


gini_index(0.1,0.9,kind=2)








gini_index(0.5,0.5,kind=3)


gini_index(0.9,0.1,kind=3)


gini_index(0.1,0.9,kind=3)

markusloecher/rfVarImpOOB documentation built on July 5, 2020, 6:50 p.m.