Search code examples
rplotclassificationrocpalette

Colorized ROC curve with cutoff values greater than one in ROCR package

I have used ROCR package to plot a colorized roc curve. The curve itself does not have any problem and looks nice and fine but the palette shows that cutoff points are greater than 1 which is incorrect as these are probabilities and should be in the range of 0 to 1. I checked my dataset several times but it seems ok and there is nothing wrong with my dataset.

Here is the code along with the predictions values and their corresponding labels.here is the plotted roc curve for my classification

My initial guess is that there is a bug in the ROCR package but I'm not completely sure. Any help to figure it out would be appreciated.

library(ROCR)
labels <-c(1,1,1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)
scores<-c(1,1,1,0.8,1,1,1,0.95,1,1,1,1,1,0,0,0,0,0,0,0.97,0,0,0,0,0,0,0,0,0,0,0,0.206,0)
pred<-prediction(scores,labels)
perf<-performance(pred,"tpr","fpr")
plot(perf,colorize=TRUE)
Solution

  • Seems you found a bug in ROCR, caused by the first cutoff being Inf, which in turn causes the first entry of alpha.values to be Inf as well:

    > attributes(pred)$cutoffs[[1]]
[1]   Inf 1.000 0.970 0.950 0.800 0.206 0.000
> attributes(perf)$alpha.values[[1]]
[1]   Inf 1.000 0.970 0.950 0.800 0.206 0.000

    This would not be a problem so far, but it seems alpha.values is used for colourization, which might cause for this strange behaviour. Setting the first entry to 1 instead of Inf fixes the issue, which could be use as a quick fix to this problem:

    > attributes(perf)$alpha.values[[1]][1] <- 1
> attributes(perf)$alpha.values[[1]]
[1] 1.000 1.000 0.970 0.950 0.800 0.206 0.000
> plot(perf,colorize=TRUE)

    ROC