2
$\begingroup$

I'm doing some work with random forests in R using the randomForest package, and I've run into something that seems odd to me. Even when the data is completely random, the AUC is never less than 0.5. For example, when I run the following:

library(randomForest)
df.sanity <- data.frame(A=sample(1:100, 2000, replace=T), B=sample(126:159, 2000, replace=T), C=sample(10:2000, 1000, replace=T), D=sample(1:2, 2000, replace=T), E=sample(30:40, 2000, replace=T), Class=as.factor(sample(0:1, 2000, replace=T)))
rf <- randomForest(x=df.sanity[1:1000,c("A", "B", "C", "D", "E")], y=df.sanity[1:1000, "Class"])
preds <- predict(rf, newdata=df.sanity[1001:2000,], type="prob")
auc(obs=df.sanity[1001:2000, "Class"], pred=preds[,2])

No matter how many times I run it, the AUC is never less than 0.5. It's often a bit over (up to 0.54 from what I've seen), but never less.

The only other AUC implementation I've used is Weka's, and I've seen AUCs < 0.5 there. Does the randomForest package automatically flip the predictions to the reverse if the AUC is ever less than 0.5, or is there something else I'm misunderstanding here?

$\endgroup$
  • $\begingroup$ That's a good point. Unfortunately, I'm at work right now on a controlled computer, so I can't install weka now. I can try when I get home. $\endgroup$ – Amanda Jul 14 '14 at 18:11
  • 1
    $\begingroup$ I am not sure whether my comment helps but once I had to write something about statistical comparison of credit default forecasts. And I also used ROC curves and calculated the AUC. I remember from the theory I read, when the two groups are independent and totally random, such that one cannot predict (whether a person belongs to 0 or 1) anything with a model,the distributions of the predicted values are almost equal. Hence, one cannot improve the model - no matter what threshold is chosen - because your ROC curve is then a straight line from 0 to 1. That is the same as saying AUC = 0.5 $\endgroup$ – random_guy Jul 14 '14 at 19:25
  • $\begingroup$ Which package is the auc() function you used (there is no auc() function in the randomForest package)? $\endgroup$ – Karl Ove Hufthammer Apr 17 '15 at 14:55
5
$\begingroup$

There is no auc() function in the randomForest package. But based on the argument names you used (obs and pred), I think you might have used the auc() function in the SDMTools package. And yes, this function does flip the results if the calculated AUC is less than 0.5:

> SDMTools::auc
function (obs, pred) 
{
    … code to calculate the AUC …
    if (AUC < 0.5) 
        AUC = 1 - AUC
    return(AUC)
}

This might be seen as a nice convenience feature (it’s easy to forget exactly how AUC functions want their arguments coded, and if you get an AUC < .5 in real life, you have usually just used the inverse/incorrect coding of the response vector ), but I think it’s a bad idea. If you try models that are (almost) as bad as random, e.g., in simulations, you will get estimates that are biased high (compared the ‘correct’ AUC estimator).

If you want the correct AUC estimates, you can use either

  • ROC() from the Epi packages. It draws a nice ROC plot, with the AUC embedded, and also returns an object with the AUC stored as the AUC element.
  • rcorr.sens() from the rms package. It returns a list with the AUC stored as the first element.
  • roc() from the pROC package if you manually specify the direction argument. It returns an object with the AUC stored as the auc element.
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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