ROC analysis widget has an option to change prior target class probability. I'd like to know, when and how it should be used. Playing with it changes the slope of iso-performance line. Each classifier should be able learn priors from the data, so I don't understand what this is for.