I have trained a classifier for 3 different classes, the test datasets of which are imbalanced, and then plotted the PR curves (below) to evaluate their accuracies. The plots contain the number of positives/negatives as well as the PR-AUC. The horizontal red line represents a baseline of the random classifier.
Similarly to the ROC that is evaluated with respect to the diagonal line, which would be produced by a completely random guess, I assume that I can also evaluate the PR curve with respect to the horizontal red line.
If this assumption is correct, can I say that the PR curves below indicate a good classification performance since they are all above the random guess line?
EDIT: figures have been updated.
Best Answer
"Good" is always subjective and problem-dependent: if the classification problem is easy, one would expect to beat the random classifier by a large margin, but for a difficult classification problem performing even just a bit better than random may be enough to be useful.