Uplift Modeling and Policy Design for Free Trial/Bonus Targeting

You ran a past randomized test that offered some users a free trial/bonus (treatment) while others received no offer (control). You want to learn who should receive the offer going forward to maximize incremental business value under budget and capacity constraints.

Design an end-to-end solution that:

1) Problem setup and labels

• Define the objective, treatment, outcome window, and the incremental metric to optimize (e.g., net profit).
• Construct training labels using only the past randomized test, clarifying intent-to-treat vs. treatment-on-the-treated.

2) Features (pre-treatment only)

• Specify pre-treatment feature groups and concrete examples. Call out any that risk leakage and how you will exclude them.

3) Modeling approach and regularization

• Choose and justify one uplift/causal modeling strategy (e.g., two-model T-learner, S-learner, X-/DR-learner, causal forests). Include base learners, cross-fitting, and regularization choices.
• Describe calibration and how you will handle class imbalance and outcome scaling (binary vs. continuous profit).

4) Offline evaluation

• Show how you would compute and interpret Qini curves/uplift AUC.
• Describe doubly-robust policy value estimation for a targeting policy.
• Explain how to incorporate treatment-capacity (top-K) and budget constraints (variable per-user costs).

5) Fairness and fraud guardrails

• Define fairness slices, metrics, and any constraints/threshold adjustments.
• Identify fraud/abuse risks (e.g., multi-accounting) and detection/mitigation tactics.

6) Online A/B policy test and safe ramp-up

• Propose an experiment design to compare the learned policy to a control/heuristic, including holdouts, interference considerations, ramp plan, and stop-loss rules.

7) Preventing leakage and handling cold-start

• List concrete steps to prevent leakage of post-treatment information in training and scoring.
• Propose approaches for cold-start users with little or no history.