Should I give treatment to all people whose predict uplift is positive or just the portion which qini curve is ascending?

[Ramlinbird]

Hello, I use the causalml's plot_qini function to draw the qini curve of my dataframe, the result is below, (The dataframe contains control/treatment random-ab-test data)

According to the curve, it seems that only the first 50% users with treatment is useful, and there is negative effect at the last.
But in the original predict uplift with treatment by the model, nearly 90% is larger than 0.
Should I chose the 90% or 30% to give the treatment? Thanks a lot.

[zeromh]

With a qini curve, you're checking the quality of your uplift model's rankings against actual uplift values. If your check with actual values shows that the last X% of your population has negative lift, then you probably don't want to target them even if your model predicts positive lift for those observations. Models aren't perfect, so this is a check.