Class flipping for uplift modeling and Heterogeneous Treatment Effect estimation on imbalanced RCT data

TL;DR

This work tackles the problem of class and treatment imbalance in uplift modeling and HTE estimation using randomized controlled trial data. It introduces class flipping as a principled alternative to undersampling, proving that flipped-class models can recover the true CATE, $\tau(x)$, or its CVT-derived counterpart without requiring calibration. The authors adapt flipping to standard classification and to CVT-based uplift, deriving conditions for identifiability and extending the method to address treatment imbalance via appropriate weighting. Empirical results on Hillstrom, Criteo, and Starbucks datasets show that Flipped CVT consistently improves uplift performance, making it a practical tool for causal inference in highly imbalanced settings.

Abstract

Uplift modeling and Heterogeneous Treatment Effect (HTE) estimation aim at predicting the causal effect of an action, such as a medical treatment or a marketing campaign on a specific individual. In this paper, we focus on data from Randomized Controlled Experiments which guarantee causal interpretation of the outcomes. Class and treatment imbalance are important problems in uplift modeling/HTE, but classical undersampling or oversampling based approaches are hard to apply in this case since they distort the predicted effect. Calibration methods have been proposed in the past, however, they do not guarantee correct predictions. In this work, we propose an approach alternative to undersampling, based on flipping the class value of selected records. We show that the proposed approach does not distort the predicted effect and does not require calibration. The method is especially useful for models based on class variable transformation (modified outcome models). We address those models separately, designing a transformation scheme which guarantees correct predictions and addresses also the problem of treatment imbalance which is especially important for those models. Experiments fully confirm our theoretical results. Additionally, we demonstrate that our method is a viable alternative also for standard classification problems.

Figures (7)

• Figure 1: Uplift curves and mAUUC for different uplift algorithms on Hillstrom dataset for logistic regression, decision tree and random forest used as base learners
• Figure 2: Uplift curves and mAUUC for different uplift algorithms on Criteo dataset for logistic regression, decision tree and random forest used as base learners
• Figure 3: Uplift curves and mAUUC for different uplift algorithms on the Starbucks dataset for logistic regression, decision tree and random forest used as base learners
• Figure 4: Effects of different class imbalance corrections classifier performance. The Method column gives the methods used to correct class imbalance: undersampling, flipping or none. For random forests $n$ denotes the number of trees in the ensemble. For forest and tree models the number below method name is the value of the 'minimum summary weight of records falling into a leaf' parameter controlling the number of records in tree leaves
• Figure 5: Uplift curves and mAUUC for different uplift algorithms on Hillstrom dataset for decision tree and random forest used as base learners