[Paper Review] Uplift Modeling for Multiple Treatments with Cost Optimization

Zhenyu Zhao, Totte Harinen|arXiv (Cornell University)|Aug 14, 2019

Advanced Causal Inference Techniques46 references4 citations

TL;DR

This paper extends meta-learner uplift models to support multiple treatment groups with varying costs, introducing net value optimization to balance conversion gains against treatment expenses. The proposed NV-Rlearner and NV-Xlearner models outperform standard methods in both net value and computation speed on synthetic and real-world data.

ABSTRACT

Uplift modeling is an emerging machine learning approach for estimating the treatment effect at an individual or subgroup level. It can be used for optimizing the performance of interventions such as marketing campaigns and product designs. Uplift modeling can be used to estimate which users are likely to benefit from a treatment and then prioritize delivering or promoting the preferred experience to those users. An important but so far neglected use case for uplift modeling is an experiment with multiple treatment groups that have different costs, such as for example when different communication channels and promotion types are tested simultaneously. In this paper, we extend standard uplift models to support multiple treatment groups with different costs. We evaluate the performance of the proposed models using both synthetic and real data. We also describe a production implementation of the approach.

Motivation & Objective

• Address the gap in uplift modeling for experiments with multiple treatment groups, which are common in industry but underexplored in existing literature.
• Integrate treatment costs into uplift modeling to enable cost-aware decision-making, as real-world interventions often have differing operational or opportunity costs.
• Evaluate the performance of extended meta-learner models in multi-treatment settings using both synthetic and real data, focusing on net value and computational efficiency.
• Develop and deploy a scalable platform implementation for online and offline uplift modeling in production environments.
• Provide empirical evidence that cost-aware models outperform standard uplift models in net value optimization, even when average conversion rates are lower.

Proposed method

• Extend two standard meta-learner frameworks—R-learner and X-learner—for multi-treatment settings by adapting their outcome and propensity score estimation to multiple treatment arms.
• Introduce a net value objective function that combines incremental conversion probability with treatment cost, enabling optimization of expected net value per user.
• Use a weighted loss function in the meta-learner training process to prioritize treatment groups with higher net value (conversion gain minus cost).
• Apply cross-validation with AUUC as the evaluation metric to select the best-performing model across different configurations.
• Implement the models in a scalable Python-based platform supporting both online real-time inference and offline batch processing.
• Integrate model outputs into production systems via APIs or batch pipelines for use in online services (e.g., in-app notifications) and offline systems (e.g., CRM platforms).

Experimental results

Research questions

• RQ1 Can standard meta-learner uplift models be effectively extended to handle multiple treatment groups beyond a single treatment and control?
• RQ2 How does incorporating treatment costs into the uplift modeling objective improve net value compared to standard uplift models that ignore cost?
• RQ3 What is the relative performance of the proposed net value-optimized models (NV-Rlearner, NV-Xlearner) versus baseline models in terms of net value and computational efficiency?
• RQ4 How do the models perform in real-world scenarios with heterogeneous treatment costs and varying conversion outcomes?
• RQ5 What are the practical challenges and design considerations for deploying uplift models at scale in production systems?

Key findings

• The NV-Rlearner and NV-Xlearner models significantly outperform standard uplift models and the original experiment groups in terms of average net value per user, with a 95% confidence interval showing statistical significance.
• Net value-optimized models achieve higher expected net value even when average conversion rates are lower than in higher-conversion treatment groups, demonstrating the value of cost-aware optimization.
• The NV-Rlearner and NV-Xlearner models showed superior performance compared to other multi-treatment uplift models in both net value and computation speed on real-world data.
• In scenarios where the incremental conversion value is lower than the treatment cost (e.g., costly phone calls), the net value-optimized models correctly avoid selecting such treatments.
• The platform implementation successfully supports both online and offline use cases, enabling real-time and batch targeting based on user-level uplift predictions.
• Empirical results show that neither standard meta-learners nor non-cost-aware models can improve net value, as they fail to account for cost trade-offs in treatment selection.