The following table shows the available methods for calculating the feature importance. Columns Classif, Regr and Surv indicate if classification, regression or survival analysis problems are supported. Columns Fac., Num. and Ord. show if a particular method can deal with factor, numeric and ordered factor features.

Current methods

Method Package Description Classif Regr Surv Fac. Num. Ord.
anova.test ANOVA Test for binary and multiclass classification tasks X X
auc AUC filter for binary classification tasks X X
carscore care CAR scores X X
FSelector_chi.squared FSelector Chi-squared statistic of independence between feature and target X X X X
FSelector_gain.ratio FSelector Chi-squared statistic of independence between feature and target X X X X
FSelector_information.gain FSelector Entropy-based information gain between feature and target X X X X
FSelector_oneR FSelector oneR association rule X X X X
FSelector_relief FSelector RELIEF algorithm X X X X
FSelector_symmetrical.uncertainty FSelector Entropy-based symmetrical uncertainty between feature and target X X X X
FSelectorRcpp_gain.ratio FSelectorRcpp Entropy-based Filters: Algorithms that find ranks of importance of discrete attributes, basing on their entropy with a continous class attribute X X X X
FSelectorRcpp_information.gain FSelectorRcpp Entropy-based Filters: Algorithms that find ranks of importance of discrete attributes, basing on their entropy with a continous class attribute X X X X
FSelectorRcpp_relief FSelectorRcpp RELIEF algorithm X X X X
FSelectorRcpp_symmetrical.uncertainty FSelectorRcpp Entropy-based Filters: Algorithms that find ranks of importance of discrete attributes, basing on their entropy with a continous class attribute X X X X
kruskal.test Kruskal Test for binary and multiclass classification tasks X X X
linear.correlation Pearson correlation between feature and target X X
mrmr mRMRe Minimum redundancy, maximum relevance filter X X X X
party_cforest.importance party Permutation importance of random forest fitted in package ‘party’ X X X X X X
permutation.importance Aggregated difference between feature permuted and unpermuted predictions X X X X X X
praznik_CMIM praznik Minimal conditional mutual information maximisation filter X X X X
praznik_DISR praznik Double input symmetrical relevance filter X X X X
praznik_JMI praznik Joint mutual information filter X X X X
praznik_JMIM praznik Minimal joint mutual information maximisation filter X X X X
praznik_MIM praznik conditional mutual information based feature selection filters X X X X
praznik_MRMR praznik Minimum redundancy maximal relevancy filter X X X X
praznik_NJMIM praznik Minimal normalised joint mutual information maximisation filter X X X X
randomForest_importance randomForest Importance based on OOB-accuracy or node inpurity of random forest fitted in package ‘randomForest’. X X X X
randomForestSRC_importance randomForestSRC Importance of random forests fitted in package ‘randomForestSRC’. Importance is calculated using argument ‘permute’. X X X X X X
randomForestSRC_var.select randomForestSRC Minimal depth of / variable hunting via method var.select on random forests fitted in package ‘randomForestSRC’. X X X X X X
ranger_impurity ranger Variable importance based on ranger impurity importance X X X X X
ranger_permutation ranger Variable importance based on ranger permutation importance X X X X X X
rank.correlation Spearman’s correlation between feature and target X X
univariate.model.score Resamples an mlr learner for each input feature individually. The resampling performance is used as filter score, with rpart as default learner. X X X X X X
variance A simple variance filter X X X X

Ensemble methods

Name Description
E-Borda Borda ensemble filter. Takes the sum across all base filter methods for each feature.
E-max Maximum ensemble filter. Takes the best maximum value across all base filter methods for each feature.
E-mean Mean ensemble filter. Takes the mean across all base filter methods for each feature.
E-median Median ensemble filter. Takes the median across all base filter methods for each feature.
E-min Minimum ensemble filter. Takes the best minimum value across all base filter methods for each feature.

Deprecated methods

Method Package Description Classif Regr Surv Fac. Num. Ord.
cforest.importance party (DEPRECATED) X X X X X X
chi.squared FSelector (DEPRECATED) X X X X
gain.ratio FSelector (DEPRECATED) X X X X
information.gain FSelector (DEPRECATED) X X X X
oneR FSelector (DEPRECATED) X X X X
randomForest.importance randomForest (DEPRECATED) X X X X
randomForestSRC.rfsrc randomForestSRC (DEPRECATED) X X X X X X
randomForestSRC.var.select randomForestSRC (DEPRECATED) X X X X X X
ranger.impurity ranger (DEPRECATED) X X X X X
ranger.permutation ranger (DEPRECATED) X X X X X X
relief FSelector (DEPRECATED) X X X X
rf.importance randomForestSRC (DEPRECATED) X X X X X X
rf.min.depth randomForestSRC (DEPRECATED) X X X X X X
symmetrical.uncertainty FSelector (DEPRECATED) X X X X
univariate (DEPRECATED) X X X X X X