Precision / Recall - Error metrics for skewed data

Some thoughts from the videos on this topic.

Please share your thoughts or comments if any gap in my understanding. Thanks.
Precision_vs_Recall - DP.pdf (78.4 KB)

1 Like

Very well written @dparekh123!

P.S. I notice a very minor thing: you used “Y >= 1” two times, did you mean “Y=1”?

Just a different interpretation to share:

So if we want to achieve the ideal situation, we will need to be able to make sure the predicted probability for actual true is always larger than that for actual false. That clear margin gives us the ideal situation, and as that clear margin disappears and a “mixed” margin shows up, we start to lose either some precision or some recall.

Yes, I meant probability threshold for Y=1. Will correct it.
Precision_vs_Recall - DP v2.pdf (78.3 KB)

Thanks Raymond for your comment.

1 Like

Yes, another way of saying let’s remove that non-overlap region of FP or FN.