I have this workflow to categorize a sample of 2000 tweets into 3 sentiment classes (POS, NEG, NEU). I'm in doubt because I get a very strange ROC curve out of the result (link to image: https://dl.dropboxusercontent.com/u/22049094/abnormal%20roc.png). Am I doing something wrong that I may not be aware of? How should I fix that? Thank you.
ROC curves are only meaningfull for two-class problems (positive, negative). You are applying it to a three-class problem. Morever, the column selection component in the dialog is for choosing one or more columns that all contain the probabilities of the positive class (e.g. different classifiers), however you have also selected columns containing probabilities of the other classes. This isn't going to work and leads to the "strange" curves for the neutral and negative classes.
Thank you for pointing that out for me :)