CODEWITHSUNDEEP
rmachine-learningdata-miningr-caret
Seyma Kalay
Seyma Kalay

Reputation: 2859

variable importance in multiclass

I have a dataset like iris and my y is a multi-class factor variable. Is there any way to see the same results for method = rf, method = treebag, and method = boost many thanks in advance.

  data(iris); head(iris)
iris$Species <- factor(iris$Species)

set.seed(87)
inTrainingSet <- createDataPartition(iris$Species, p=.80,list=0)
train <- iris[inTrainingSet,]
test  <- iris[-inTrainingSet,]
ctrl <- trainControl(method = "cv", number = 2,  verboseIter = TRUE)


pls <- train(Species ~ Sepal.Length+Sepal.Width+Petal.Length+Petal.Width , 
              method = "pls", data = iris,
              trControl = ctrl)
attributes(varImp(pls))
varImp(pls)$importance

Upvotes: 2

Views: 869

Answers (1)

StupidWolf
StupidWolf

Reputation: 46898

There's a few points to your question, so if there is a built in method to estimate this properly for each model, you can run varImp with the default useModel = FALSE.

For randomforest, you add importance=TRUE while fitting:

rf <- train(Species ~ Sepal.Length+Sepal.Width+Petal.Length+Petal.Width , 
              method = "rf", data = iris,
              trControl = ctrl,importance=TRUE)
varImp(rf)

rf variable importance

  variables are sorted by maximum importance across the classes
             setosa versicolor virginica
Petal.Length  66.94     100.00     85.40
Petal.Width   63.86      92.22     89.87
Sepal.Length  16.75      24.05     24.90
Sepal.Width   12.75       0.00     17.49

If the model does not have an inbuilt for multiclass, then the pairwise roc curve is used to derive these importances, see page for caret on the specificities on this:

tb <- train(Species ~ Sepal.Length+Sepal.Width+Petal.Length+Petal.Width , 
                  method = "treebag", data = iris,
                  trControl = ctrl,importance=TRUE)

varImp(tb,useModel=TRUE)
treebag variable importance

             Overall
Petal.Length  100.00
Petal.Width    99.17
Sepal.Length   32.23
Sepal.Width     0.00

 varImp(tb,useModel=FALSE)
ROC curve variable importance

  variables are sorted by maximum importance across the classes
             setosa versicolor virginica
Petal.Width  100.00     100.00     100.0
Petal.Length 100.00     100.00     100.0
Sepal.Length  90.70      59.30      90.7
Sepal.Width   54.59      54.59       0.0

You did not specify which boosted tree method use, but I guess you can easily use one of the options above

Upvotes: 2

Related Questions