Great video. Thanks! So if PR seems to work for both balanced and imbalanced data sets, why would you not just always use PR curves? When would ROC make more sense?
If I have a distance metric as the output of a model ( say euclidean distance in face verification for matching and mismatched pairs). How do you choose a cut off of the euclidean distance ? I guess we can use same concept only a low score is indicative of +ve match class and high score is indicative of a -ve mismatch true negative class
one technique i did was to divide the eulidean distances by 100 ( so 15.37 for a mismatch would be .1537, and 3.23 for a match case would be .0323, then i would subtract it from 1 so that they look like probabilites of similarity , can i then use these to plot the ROC curves ? SO that i can choose a threshold with high TPR and low FPR.
Bro, this is a very good explanation. This information is rare. Thanks a lot
Great video. Thanks! So if PR seems to work for both balanced and imbalanced data sets, why would you not just always use PR curves? When would ROC make more sense?
We have been waiting for THIS!!!!! Thank you
I've been looking at ROC AUC score for my unbalanced dataset. I will have to look at PR AUC instead, thank you.
this helped me so much with my unbalanced data.
Super glad it did!
Haha this is *exactly* what I was looking for (implementing the curves from scratch).
Thanks mate!
I know it's quite off topic but do anybody know of a good website to watch new movies online ?
@Ivan Allan I would suggest Flixzone. Just google for it :)
@Maddux Sam Definitely, have been watching on flixzone for since april myself =)
@Maddux Sam thank you, signed up and it seems like they got a lot of movies there :) I really appreciate it !!
@Ivan Allan happy to help xD
these coding videos are lit
Thanks so much, definitely needed this
thank you so much
yet another awesome video. your amazing
one advice from me. please change your profile photo it makes your channel seems less professional.
If I have a distance metric as the output of a model ( say euclidean distance in face verification for matching and mismatched pairs). How do you choose a cut off of the euclidean distance ? I guess we can use same concept only a low score is indicative of +ve match class and high score is indicative of a -ve mismatch true negative class
one technique i did was to divide the eulidean distances by 100 ( so 15.37 for a mismatch would be .1537, and 3.23 for a match case would be .0323, then i would subtract it from 1 so that they look like probabilites of similarity , can i then use these to plot the ROC curves ? SO that i can choose a threshold with high TPR and low FPR.