Rethinking Importance Weighting for Deep Learning under Distribution Shift T ongtong Fang 1 Nan Lu1,2 Gang Niu

Neural Information Processing Systems 

Eq. (1), importance weighting (IW) [46, 49, 21, 50, 51, 25] can handle DS in two separate steps: weight estimation (WE) with the help of a tiny set of validation data from p