My model is working ok (the AUC is 0.7) but the importances from a randomForest run for my binary classification problem differ depending on how I retrieve them. Is this normal? They seem to be scaled up when I call importance(rf) (by order 105). More importantly, the first three columns don't match at all, even discounting the magnitude! I'm unsure how to interpret this at the moment having read ?importance and ?randomForest.

Here's my randomForest call:

rf <- randomForest(trnX, as.factor(trnY), ntree=1000, importance=TRUE)

and here's a portion of the outputs:


                   0             1 MeanDecreaseAccuracy MeanDecreaseGini
var1   -2.308793e-05 -6.124117e-05        -3.642557e-05         30.47050
var2    6.169346e-04 -3.947637e-04         2.563570e-04         48.04932
var3    3.621287e-03  1.834355e-03         2.981152e-03         68.28302
var4    3.655234e-03  1.981978e-03         3.057267e-03         79.12254
var5    5.350649e-03  1.041555e-03         3.812376e-03         84.88832
var6    3.366199e-03  1.707144e-03         2.773894e-03         78.02293


                0          1 MeanDecreaseAccuracy MeanDecreaseGini
var1   -0.4739717 -0.8497967           -0.8952541         30.47050
var2    7.9028043 -3.6680758            4.1625329         48.04932
var3   17.1279919  5.9733456           20.8354079         68.28302
var4   17.3895593  7.2370419           22.4948931         79.12254
var5   20.7470555  3.2702612           24.3069376         84.88832
var6   16.9071520  6.2532409           21.1648686         78.02293

My data has 8000 data points and 60 variables. There is definitely correlation between variables and they are of different types: some are binary flags, other are integers, others continuous. Some variables have missing values (which I'm fixing with na.roughfix()).

Additionally, a run from gbm:

gbmBernModel <- gbm(formula = var1 ~ ., distribution="bernoulli", 
                    data = trn, verbose="CV", n.trees=10000, interaction.depth=2)

produced relative importances (viewed by summary(gbmBernModel)) that made logical sense to me.

Any pointers would be greatly appreciated. Thank you.


1 Answer 1


The default method of importance() is after scaling the importance "value", i.e. for permutation based measures, the measures have been divided their “standard errors”. While the rf$importance is not.

So if you try the code of

importance(rf, scale=FALSE)

The result is the same as rf$importance

  • $\begingroup$ Thanks @Vincent. From this it follows that importance(rf) is equivalent to rf$importance/rf$importanceSD $\endgroup$ Commented Nov 18, 2013 at 17:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.