Technical Summary: Guardrailed Uplift Targeting

Problem Statement

Modern e-commerce platforms face a critical challenge in optimizing customer targeting strategies while simultaneously adhering to strict business constraints. Traditional approaches, such as propensity scoring or random assignment, often fail to account for heterogeneous treatment effects (HTE) and business limitations concurrently. This leads to suboptimal resource allocation, where investments in promotions or retention campaigns may yield negative returns or even harm customer experience (e.g., by triggering reactance in high-risk segments). The core problem addressed is how to translate causal estimates of treatment effects into actionable targeting policies that systematically respect diverse constraints, including budget limits, revenue protection, and fairness metrics.

Methodological Framework

The paper proposes a two-stage "Guardrailed Uplift Targeting" framework that integrates causal machine learning with constrained optimization.

1. Problem Formulation

The objective is to learn a targeting policy $\pi: \mathcal{X} \to \mathcal{T}$ that maximizes expected outcomes $E[Y(\pi(X))]$ subject to business constraints $g_j(\pi) \leq c_j$. The policy assigns exactly one treatment (including a control/baseline) to each customer from a set of available treatments.

2. Stage I: Uplift Estimation (CATE)

The framework first estimates Conditional Average Treatment Effects (CATE) using state-of-the-art uplift learners. The CATE for a customer $x_i$ is defined as the difference in potential outcomes between treatment and control:
$$\tau(x_i) = E[Y_i(1)|X_i=x_i] - E[Y_i(0)|X_i=x_i]$$
The authors employ several estimators, including:

• Causal Forest: Adapts random forests for high-dimensional covariates.
• Double Machine Learning (DML): A doubly robust method combining outcome and propensity score modeling.
• Forest Doubly Robust Learner: Integrates forest methods with doubly robust estimation.

3. Stage II: Constrained Optimization

Using the estimated uplifts $\hat{\tau}_k(x_i)$, the framework solves a constrained optimization problem to determine the optimal assignment matrix $\pi$:
$$\max_{\pi} \sum_{i=1}^{N} \sum_{k=1}^{K} \pi_{ik} w_i \hat{\tau}_k(x_i)$$
Subject to:

• Binary Assignment: $\pi_{ik} \in \{0, 1\}$
• Single Treatment: $\sum_k \pi_{ik} = 1$
• Business Guardrails: $g(\pi) \leq c$

Constraint Types:

• Budget: Limiting the proportion of customers receiving a specific treatment.
• Revenue Protection: Ensuring total revenue does not drop below a threshold (e.g., 98% of the "treat-all" scenario).
• Fairness: Enforcing demographic parity or equal opportunity.
• Operational Limits: Capacity constraints on delivery systems.

To address computational scalability for large populations, the framework utilizes customer bucketing, grouping customers by similar uplift estimates and optimizing over groups rather than individuals.

4. Evaluation Methodology

Performance is assessed via:

• Uplift Model Metrics: Area Under the Cumulative Uplift Curve (AUC).
• Offline Policy Evaluation: Inverse Propensity Scoring (IPS) and Self-Normalized IPS (SNIPS) to estimate policy value from historical data.
• Online Validation: Large-scale A/B testing against business-as-usual baselines.

Key Contributions

The paper outlines five primary contributions:

1. A Two-Stage Framework: A novel integration of CATE estimation with constrained optimization specifically designed for practical marketing decisions.
2. Simulation Setup: A detailed offline validation environment including data generation, treatment assignment, response functions, and specific guardrail constraints.
3. Empirical Validation: Demonstrations across three distinct business scenarios (Retention, Event Revenue, Spend Thresholds) showing consistent improvements over baselines.
4. Implementation Guidelines: Practical advice on data requirements, model selection, constraint specification, and computational scalability.
5. Open Resources: Provision of publicly available code and a reusable playbook for operationalizing causal targeting.

Experimental Results

The framework was validated through offline simulations and online A/B tests across three scenarios:

A. Customer Retention Optimization

• Context: Reducing churn via proactive messaging.
• Finding: The optimized policy targeted only 6.47% of customers (vs. ~34% for the baseline).
• Outcome: Achieved a 2.35% relative lift in retention rates compared to the propensity-based baseline. Crucially, it identified that customers with very low retention scores responded negatively to messaging and excluded them, avoiding the "treat-all" inefficiency.

B. Event Revenue Maximization

• Context: Assigning tiered rewards (P1 vs. P2) for promotional events.
• Constraint: Revenue protection (sales must not drop >1% compared to assigning the max reward to all).
• Outcome: The Causal Forest DML model achieved a 0.317% lift in revenue and reduced incentive spend as a percentage of sales by 11.879%. It also improved completion rates by 1.02%, demonstrating that cost efficiency did not come at the expense of customer experience.

C. Spend Threshold Optimization

• Context: Assigning optimal spend thresholds (S1 vs. S2) to qualify for rewards.
• Outcome: In a large-scale A/B test, the optimized policy achieved a statistically significant 0.36% revenue lift ($p=0.040$) and a 5.49% improvement in completion rate ($p=0.000$). The policy assigned the lower threshold to only 11.93% of customers (vs. 34.49% for the baseline), proving that selective targeting drives better overall performance.

Significance and Claims

The paper claims that this framework provides a systematic methodology for converting causal estimates into executable marketing strategies that respect real-world business limitations. By moving beyond simple propensity scoring, the approach enables marketers to:

• Identify optimal customer segments that are truly responsive to interventions.
• Avoid targeting segments that may react negatively (e.g., high-risk churners).
• Operate within strict budget and revenue guardrails, ensuring that optimization does not compromise core business metrics.

The authors position this work as a reusable playbook for scaling causal targeting in e-commerce, validated by both offline simulations and online A/B tests that consistently outperform static baselines and propensity-based approaches. The framework bridges the gap between advanced causal machine learning theory and practical deployment requirements, offering a path to more efficient, ethical, and profitable marketing operations.