Imbalance-Aware Uplift Modeling for Observational Data

Uplift modeling aims to model the incremental impact of a treatment on an individual outcome, which has attracted great interests researchers and practitioners from different communities. Existing uplift methods rely either data collected randomized controlled trials (RCTs) or observational is more realistic. However, we notice that data, it often case only small number subjects receive treatment, but finally infer much large group subjects. Such highly imbalanced common in various fields such as marketing medical rarely handled by existing works. In this paper, theoretically quantitatively prove representative methods, transformed outcome (TOM) doubly robust (DR), suffer bias deviation datasets with skewed propensity scores, mainly because they are proportional reciprocal score. To reduce dataset, propose imbalance-aware (IAUM) method via constructing proxy adaptively combines estimator imputed effects based We IAUM can obtain better bias-variance trade-off than dataset. conduct extensive experiments synthetic dataset two real-world datasets, experimental results well demonstrate superiority our over state-of-the-art.