Build predictive model for feature rollout targeting

Before global launch, you want to predict which users or products would benefit most from the 'More like this' button so you can stage rollout. Design an end-to-end modeling approach using only pre-launch data from the interactions/products schema and any new logging you can add at launch:

1. Labeling: propose a proxy label available pre-launch (e.g., user propensity to explore similar items via existing flows) and a post-launch true label (uplift in exploration rate or interaction_count per user). Explain how you will avoid target leakage, especially from features derived too close to the labeling window.
2. Features: list concrete user, product, and interaction features you’ll engineer (e.g., recency/frequency of category interactions, dwell-time proxies via interaction_count sequences, country, seller diversity). Include cross-features (user×category affinity) and cold-start strategies.
3. Models: choose two candidate model families (e.g., calibrated gradient-boosted trees and a sparse logistic regression). Explain when a deep model would be justified (sample size, feature types) and how you’d ensure calibration for decisioning.
4. Evaluation: define offline splits (time-based, user-level holdout), metrics (AUC/PR for ranking; Qini/uplift AUC if modeling treatment uplift), and how you’ll run prospective shadow testing once telemetry exists.
5. Decisioning: specify a thresholding or budgeted targeting policy and show how you’d simulate business impact under deployment constraints. Address fairness across countries and guard against negative spillovers (e.g., reduced variety). Include a plan for periodic retraining and drift monitoring.