Type: Package
Title: Intervention in Prediction Measure for Random Forests
Version: 1.3
Date: 2025-09-13
Maintainer: Irene Epifanio <epifanio@uji.es>
Imports: party, randomForest, gbm
Suggests: mlbench
Description: Computes intervention in prediction measure for assessing variable importance for random forests. See details at I. Epifanio (2017) <doi:10.1186/s12859-017-1650-8>.
License: GPL-3
NeedsCompilation: no
Packaged: 2025-09-13 09:55:45 UTC; Vicente
Author: Irene Epifanio [aut, cre], Stefano Nembrini [aut], Aleix Alcacer [aut], Vicente Bolos [aut]
Repository: CRAN
Date/Publication: 2025-09-29 08:20:02 UTC

Intervention in Prediction Measure (IPM) for Random Forests

Description

It computes IPM for assessing variable importance for random forests. See I. Epifanio (2017). Intervention in prediction measure: a new approach to assessing variable importance for random forests. BMC Bioinformatics.

Details

Package: IPMRF
Type: Package
Version: 1.2
Date: 2017-08-09

Main Functions:

Author(s)

Irene Epifanio, Stefano Nembrini

References

Pierola, A. and Epifanio, I. and Alemany, S. (2016) An ensemble of ordered logistic regression and random forest for child garment size matching. Computers & Industrial Engineering, 101, 455–465.

Epifanio, I. (2017) Intervention in prediction measure: a new approach to assessing variable importance for random forests. BMC Bioinformatics, 18, 230.

See Also

https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-017-1650-8

Internal function(s)

Description

They are internal functions used for returning which variables are used in prediction in tree with randomForest and party.

Value

An array with the results

Author(s)

Irene Epifanio

Internal function for gbm

Description

It is an internal function used for reading trees from gbm with gbm.

Value

A data frame with the results

Author(s)

Stefano Nembrini

IPM casewise with gbm object by gbm for new cases, whose responses do not need to be known

Description

The IPM of a new case, i.e. one not used to grow the forest and whose true response does not need to be known, is computed as follows. The new case is put down each of the ntree trees in the forest. For each tree, the case goes from the root node to a leaf through a series of nodes. The variable split in these nodes is recorded. The percentage of times a variable is selected along the case's way from the root to the terminal node is calculated for each tree. Note that we do not count the percentage of times a split occurred on variable k in tree t, but only the variables that intervened in the prediction of the case. The IPM for this new case is obtained by averaging those percentages over the ntree trees.

Usage

ipmgbmnew(marbolr, da, ntree)

Arguments

marbolr

Generalized Boosted Regression object obtained with gbm.

da

Data frame with the predictors only, not responses, for the new cases. Each row corresponds to an observation and each column corresponds to a predictor, which obviously must be the same variables used as predictors in the training set.

ntree

Number of trees.

Details

All details are given in Epifanio (2017).

Value

It returns IPM for new cases. It is a matrix with as many rows as cases are in da, and as many columns as predictors are in da.

Note

See Epifanio (2017) about the parameters of RFs to be used, the advantages and limitations of IPM, and in particular when CART is considered with predictors of different types.

Author(s)

Stefano Nembrini

References

Pierola, A. and Epifanio, I. and Alemany, S. (2016) An ensemble of ordered logistic regression and random forest for child garment size matching. Computers & Industrial Engineering, 101, 455–465.

Epifanio, I. (2017) Intervention in prediction measure: a new approach to assessing variable importance for random forests. BMC Bioinformatics, 18, 230.

See Also

ipmparty, ipmrf, ipmpartynew, ipmrfnew

Examples


library(party)
library(gbm)
gbm = gbm(score ~ ., data = readingSkills, n.trees = 50, shrinkage = 0.05, interaction.depth = 5, 
        bag.fraction = 0.5, train.fraction = 0.5, n.minobsinnode = 1, 
        cv.folds = 0, keep.data = FALSE, verbose = FALSE)
apply(ipmgbmnew(gbm, readingSkills[, -4], 50), FUN = mean, 2) -> gbm_ipm
gbm_ipm

IPM casewise with CIT-RF by party for OOB samples

Description

The IPM for a case in the training set is calculated by considering and averaging over only the trees where the case belongs to the OOB set. The case is put down each of the trees where the case belongs to the OOB set. For each tree, the case goes from the root node to a leaf through a series of nodes. The variable split in these nodes is recorded. The percentage of times a variable is selected along the case's way from the root to the terminal node is calculated for each tree. Note that we do not count the percentage of times a split occurred on variable k in tree t, but only the variables that intervened in the prediction of the case. The IPM for this case is obtained by averaging those percentages over only the trees where the case belongs to the OOB set. The random forest is based on CIT (Conditional Inference Trees).

Usage

ipmparty(marbol, da, ntree)

Arguments

marbol

Random forest obtained with cforest . Responses can be of the same type supported by cforest, not only numerical or nominal, but also ordered responses, censored response variables and multivariate responses.

da

Data frame with the predictors only, not responses, of the training set used for computing marbol. Each row corresponds to an observation and each column corresponds to a predictor. Predictors can be numeric, nominal or an ordered factor.

ntree

Number of trees in the random forest.

Details

All details are given in Epifanio (2017).

Value

It returns IPM for cases in the training set. It is estimated when they are OOB observations. It is a matrix with as many rows as cases are in da, and as many columns as predictors are in da. IPM can be estimated for any kind of RF computed by cforest, including multivariate RF.

Note

See Epifanio (2017) about advantages and limitations of IPM, and about the parameters to be used in cforest.

Author(s)

Irene Epifanio

References

Pierola, A. and Epifanio, I. and Alemany, S. (2016) An ensemble of ordered logistic regression and random forest for child garment size matching. Computers & Industrial Engineering, 101, 455–465.

Epifanio, I. (2017) Intervention in prediction measure: a new approach to assessing variable importance for random forests. BMC Bioinformatics, 18, 230.

See Also

ipmpartynew, ipmrf, ipmrfnew, ipmgbmnew

Examples

# Note: more examples can be found at 
# https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-017-1650-8

## -------------------------------------------------------
## Example from \code{\link[party]{varimp}} in \pkg{party}
## Classification RF
## -------------------------------------------------------


library(party)

# from help in varimp by party package
set.seed(290875)
readingSkills.cf <- cforest(score ~ ., data = readingSkills,
control = cforest_unbiased(mtry = 2, ntree = 50))

# standard importance
varimp(readingSkills.cf)

# the same modulo random variation
varimp(readingSkills.cf, pre1.0_0 = TRUE)

# conditional importance, may take a while...
varimp(readingSkills.cf, conditional = TRUE)


# IMP based on CIT-RF (party package)
library(party)

ntree <- 50
# readingSkills: data from party package
da <- readingSkills[, 1:3] 
set.seed(290875)
readingSkills.cf3 <- cforest(score ~ ., data = readingSkills,
control = cforest_unbiased(mtry = 3, ntree = 50))

# IPM case-wise computed with OOB with party
pupf <- ipmparty(readingSkills.cf3, da, ntree)
 
# global IPM
pua <- apply(pupf, 2, mean) 
pua

IPM casewise with CIT-RF by party for new cases, whose responses do not need to be known

Description

The IPM of a new case, i.e. one not used to grow the forest and whose true response does not need to be known, is computed as follows. The new case is put down each of the ntree trees in the forest. For each tree, the case goes from the root node to a leaf through a series of nodes. The variable split in these nodes is recorded. The percentage of times a variable is selected along the case's way from the root to the terminal node is calculated for each tree. Note that we do not count the percentage of times a split occurred on variable k in tree t, but only the variables that intervened in the prediction of the case. The IPM for this new case is obtained by averaging those percentages over the ntree trees. The random forest is based on CIT (Conditional Inference Trees).

Usage

ipmpartynew(marbol, da, ntree)

Arguments

marbol

Random forest obtained with cforest . Responses in the training set can be of the same type supported by cforest, not only numerical or nominal, but also ordered responses, censored response variables and multivariate responses.

da

Data frame with the predictors only, not responses, for the new cases. Each row corresponds to an observation and each column corresponds to a predictor, which obviously must be the same variables used as predictors in the training set. Predictors can be numeric, nominal or an ordered factor.

ntree

Number of trees in the random forest.

Details

All details are given in Epifanio (2017).

Value

It returns IPM for new cases. It is a matrix with as many rows as cases are in da, and as many columns as predictors are in da. IPM can be estimated for any kind of RF computed by cforest, including multivariate RF.

Note

See Epifanio (2017) about advantages and limitations of IPM, and about the parameters to be used in cforest.

Author(s)

Irene Epifanio

References

Pierola, A. and Epifanio, I. and Alemany, S. (2016) An ensemble of ordered logistic regression and random forest for child garment size matching. Computers & Industrial Engineering, 101, 455–465.

Epifanio, I. (2017) Intervention in prediction measure: a new approach to assessing variable importance for random forests. BMC Bioinformatics, 18, 230.

See Also

ipmparty, ipmrf, ipmrfnew, ipmgbmnew

Examples

# Note: more examples can be found at 
# https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-017-1650-8

## -------------------------------------------------------
## Example from \code{\link[party]{varimp}} in \pkg{party}
## Classification RF
## -------------------------------------------------------

library(party)

# IMP based on CIT-RF (party package)
ntree = 50
# readingSkills: data from party package
da = readingSkills[,1:3] 
set.seed(290875)
readingSkills.cf3 <- cforest(score ~ ., data = readingSkills,
control = cforest_unbiased(mtry = 3, ntree = 50))

# new case
nativeSpeaker = 'yes'
age = 8
shoeSize = 28
da1 = data.frame(nativeSpeaker, age, shoeSize)

# IPM case-wise computed for new cases for party package
pupfn = ipmpartynew(readingSkills.cf3, da1, ntree)
pupfn

IPM casewise with CART-RF by randomForest for OOB samples

Description

The IPM for a case in the training set is calculated by considering and averaging over only the trees where the case belongs to the OOB set. The case is put down each of the trees where the case belongs to the OOB set. For each tree, the case goes from the root node to a leaf through a series of nodes. The variable split in these nodes is recorded. The percentage of times a variable is selected along the case's way from the root to the terminal node is calculated for each tree. Note that we do not count the percentage of times a split occurred on variable k in tree t, but only the variables that intervened in the prediction of the case. The IPM for this case is obtained by averaging those percentages over only the trees where the case belongs to the OOB set. The random forest is based on CART.

Usage

ipmrf(marbolr, da, ntree)

Arguments

marbolr

Random forest obtained with randomForest. Responses can be of the same type supported by randomForest. Note that not only numerical or nominal, but also ordered responses, censored response variables and multivariate responses can be considered with ipmparty.

da

Data frame with the predictors only, not responses, of the training set used for computing marbolr. Each row corresponds to an observation and each column corresponds to a predictor. Predictors can be numeric, nominal or an ordered factor.

ntree

Number of trees in the random forest.

Details

All details are given in Epifanio (2017).

Value

It returns IPM for cases in the training set. It is estimated when they are OOB observations. It is a matrix with as many rows as cases are in da, and as many columns as predictors are in da.

Note

See Epifanio (2017) about the parameters of RFs to be used, the advantages and limitations of IPM, and in particular when CART is considered with predictors of different types.

Author(s)

Irene Epifanio

References

Pierola, A. and Epifanio, I. and Alemany, S. (2016) An ensemble of ordered logistic regression and random forest for child garment size matching. Computers & Industrial Engineering, 101, 455–465.

Epifanio, I. (2017) Intervention in prediction measure: a new approach to assessing variable importance for random forests. BMC Bioinformatics, 18, 230.

See Also

ipmparty, ipmpartynew, ipmrfnew, ipmgbmnew

Examples

# Note: more examples can be found at 
# https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-017-1650-8



library(mlbench)
# data used by Breiman, L.: Random forests. Machine Learning 45(1), 5--32 (2001)
data(PimaIndiansDiabetes2) 
Diabetes <- na.omit(PimaIndiansDiabetes2)

set.seed(2016)
require(randomForest)
ri<- randomForest(diabetes  ~ ., data = Diabetes, ntree = 500, importance = TRUE,
keep.inbag = TRUE, replace = FALSE) 

# GVIM and PVIM (CART-RF)
im = importance(ri)
im
# rank
ii = apply(im, 2, rank)
ii

# IPM based on CART-RF (randomForest package)
da = Diabetes[, 1:8]
ntree = 500
# IPM case-wise computed with OOB 
pupf = ipmrf(ri, da, ntree) 

# global IPM
pua = apply(pupf, 2, mean) 
pua

# IPM by classes
attach(Diabetes)
puac = matrix(0, nrow = 2, ncol = dim(da)[2])
puac[1, ] = apply(pupf[diabetes == 'neg', ], 2, mean) 
puac[2, ] = apply(pupf[diabetes == 'pos', ], 2, mean) 
colnames(puac) = colnames(da)
rownames(puac) = c('neg', 'pos')
puac

# rank IPM 
# global rank 
rank(pua) 
# rank by class
apply(puac, 1, rank)

IPM casewise with CART-RF by randomForest for new cases, whose responses do not need to be known

Description

The IPM of a new case, i.e. one not used to grow the forest and whose true response does not need to be known, is computed as follows. The new case is put down each of the ntree trees in the forest. For each tree, the case goes from the root node to a leaf through a series of nodes. The variable split in these nodes is recorded. The percentage of times a variable is selected along the case's way from the root to the terminal node is calculated for each tree. Note that we do not count the percentage of times a split occurred on variable k in tree t, but only the variables that intervened in the prediction of the case. The IPM for this new case is obtained by averaging those percentages over the ntree trees.

The random forest is based on CART

Usage

ipmrfnew(marbolr, da, ntree)

Arguments

marbolr

Random forest obtained with randomForest. Responses can be of the same type supported by randomForest. Note that not only numerical or nominal, but also ordered responses, censored response variables and multivariate responses can be considered with ipmparty.

da

Data frame with the predictors only, not responses, for the new cases. Each row corresponds to an observation and each column corresponds to a predictor, which obviously must be the same variables used as predictors in the training set.

ntree

Number of trees in the random forest.

Details

All details are given in Epifanio (2017).

Value

It returns IPM for new cases. It is a matrix with as many rows as cases are in da, and as many columns as predictors are in da.

Note

See Epifanio (2017) about the parameters of RFs to be used, the advantages and limitations of IPM, and in particular when CART is considered with predictors of different types.

Author(s)

Irene Epifanio

References

Pierola, A. and Epifanio, I. and Alemany, S. (2016) An ensemble of ordered logistic regression and random forest for child garment size matching. Computers & Industrial Engineering, 101, 455–465.

Epifanio, I. (2017) Intervention in prediction measure: a new approach to assessing variable importance for random forests. BMC Bioinformatics, 18, 230.

See Also

ipmparty, ipmrf, ipmpartynew, ipmgbmnew

Examples

#Note: more examples can be found at 
#https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-017-1650-8


library(mlbench)
# data used by Breiman, L.: Random forests. Machine Learning 45(1), 5--32 (2001)
data(PimaIndiansDiabetes2) 
Diabetes <- na.omit(PimaIndiansDiabetes2)

set.seed(2016)
require(randomForest)
ri <- randomForest(diabetes  ~ ., data = Diabetes, ntree = 500, importance = TRUE,
keep.inbag = TRUE, replace = FALSE) 

# new cases
da1 = rbind(apply(Diabetes[Diabetes[, 9] == 'pos', 1:8], 2, mean),
apply(Diabetes[Diabetes[, 9] == 'neg', 1:8], 2, mean))


# IPM case-wise computed for new cases for randomForest package
ntree = 500
pupfn = ipmrfnew(ri, as.data.frame(da1), ntree)
pupfn