jayfjin / 10-701-project Goto Github PK

There's another interesting aspect we can look at that's not covered in the paper. They looked at the F1 score, which penalizes false positives and false negatives equally. But in practice, false positives are much worse, because we can show a little bit of a commercial as long as we don't cut out any of the actual show. The F1 score can be modified to weight false positives more heavily, so we can use that in the evaluation too, and consider a learning algorithm that can use a custom loss function that has a larger penalty for false positives.