featureContributions: Feature Contributions for a Random Forest Model

In rfFC: Random Forest Feature Contributions

Description

This method calculates feature contributions for a given dataset and an existing Random Forest model randomForest. The feature contributions are computed separately for each instance/record in dataset and provide detailed information about relationships between variables and the predicted value. This method was implemented based upon the approach of Kuz'min et al. for regression models and extended to classification models. For a binary classification model the method returns the feature contributions towards class "one". For a multi-class model, the feature contributions are calculated towards the class predicted by the randomForest model for a given instance.

The method does not work for unsupervised models. The randomForest model must have a stored in-bag matrix that keeps track of which samples were used to build trees in the forest and sampling without replacement must be used to generate a model.

Hence, all Random Forest models analyzed by this method must be generated as follows:

model <- randomForest(...,keep.inbag=TRUE,replace=FALSE)

The reason for this current limitation is because, in the code of the randomForest implementation of Random Forest provided by Liaw and Wiener, the inbag matrix does not record how many times a sample was used to build a particular tree (if sampling with replacement).

Usage

featureContributions(object, lInc, dataT, mClass=NULL)

Arguments

object	an object of the class randomForest
lInc	local increments of feature contributions calculated for this object using getLocalIncrements
dataT	a data frame containing the variables in the model (columns) for all instances (rows) for which feature contributions are desired
mClass	a name of the class to which feature contributions is calculated. The class name must to match to the one class name from the randomForest object variable y. By default, the value of this parameter is set to NULL. In this case the feature contributions are calculated to the predicted class returned by the predict method from the randomForest package. This option is avaliable only for the multiclassification problems.

Value

A list with the following components:

contrib

n x m matrix of feature contributions, where n is the number of records (i.e. instances) and m is the number of features/variables (i.e. attributes/descriptors) of the dataset dataT. For regression, or multi-class classification, the feature contributions represent the signed contributions towards the predicted value or a given class defined by the argument mClass. For binary classification, by default, the classes are internally treated as numeric 1 and 0 and the feature contributions represent the signed contributions towards "class 1". If the class labels in the training set are presented as "1" and "0", then the corresponding classes will be internally treated as 1 and 0 respectively; otherwise, this mapping will be performed arbitrarily with the class of the first instance in the training set treated internally as "class 1", or to the class provided as a parameter in getLocalIncrements.

Author(s)

Anna Palczewska [e-mail] and
Richard Marchese Robinson [e-mail]

References

V.E. Kuz'min et al. (2011), Interpretation of QSAR Models Based on Random Forest Methods, Molecular Informatics, 30, 593-603.
A. Palczewska et al. (2013), Interpreting random forest models using a feature contribution method, Proceedings of the 2013 IEEE 14th International Conference on Information Reuse and Integration IEEE IRI 2013, August 14-16, 2013, San Francisco, California, USA, 112-119.
A. Palczewska et al. (2014), Interpreting random forest classification models using a feature contribution method. in Integration of Reusable Systems, ser. Advances in Intelligent and Soft Computing, T. Bouabana-Tebibel and S. H. Rubin, Eds. Springer International Publishing, 263, 193-218.

See Also

randomForest, getLocalIncrements

Examples

#Multi-class Classification
library(randomForest)
data(iris)
rF <- randomForest(x=iris[,-5],y=as.factor(as.character(iris[,5])),
    ntree=25,importance=TRUE, keep.inbag=TRUE,replace=FALSE) 
#Get Local feature incremets    
li<-getLocalIncrements(rF, iris[,-5])
#Calculate feature contributions
fc<-featureContributions(rF, li, iris[,-5])    

## Not run: 
#Binary classification
library(randomForest)
data(ames)
ames_train<-ames[ames$Type=="Train",-c(1,3, ncol(ames))]
rF_Model <- randomForest(x=ames_train[,-1],y=as.factor(as.character(ames_train[,1])),
          ntree=500,importance=TRUE, keep.inbag=TRUE,replace=FALSE) 
li <- getLocalIncrements(rF_Model,ames_train[,-1])
fc<-featureContributions(rF_Model, li, ames_train[,-1])

## End(Not run)

rfFC documentation built on May 2, 2019, 5:18 p.m.