So far, my biggest claim to fame on twitter has been my animations about receiver operating characteristic (ROC) curves. You may have seen them circling around the twitterverse.
I finally got around to trying #gganimate. I made this visualization for the bioinformatics class @ClausWilke and I are teaching next semester. I am hoping this will help students understand how ROC curves behave for different linear predictors. #rstats pic.twitter.com/SXMdLmvKR8— Dariya Sydykova (@dariyasydykova) November 14, 2018
AUC can be misleading when the std dev of one of the outcomes decreases. Even though AUC increases, the prediction performance is worse at small false positive rates. #gganimate #rstats pic.twitter.com/YzMOH0KX8V— Dariya Sydykova (@dariyasydykova) December 14, 2018
Now a disclaimer: I have never made one ROC or a precision-recall curve for my research. Everything I learned about these curves I have learned from from making these animations and from getting feedback from others that are more familiar with them. As such, I figured I should write a blogpost that would be a solid primer to get someone entirely new to the concept to get started on understanding them.
So here it is
Classifying binary outcome
ROC and precision-recall curves evaluate how well a binary classifier