myazdani / regularized-t-learner

Investigating T-learners with regularization

regularized-t-learner

Investigating T-learners with regularization

What's a T-learner? The "T" is for two, meaning a T-learner is actually two learners.

What's a T-learner learning? The uplift.

What's the uplift? The difference between two binary classifications under two independent data regimes. Specifically, Uplift = P(y|x,T=1) - p(y|x,T=0)

For example, usually in binarly classification we are just learning P(y|x), say what are the chances a customer x makes purchase y. But if we want to know how the chance of a purchase changes when showing a coupon, we need to measure the uplift between showing and not showing a coupon. From a random and representative population, we collect data T = 1 where a coupon is shown to customers and see how many make purchases. Then from another random and representative population, we collect data T = 0 where a coupon is not shown. We learn two separate binary classifiers with two different datasets (T=0, x, y) and (T=1, x, y).