Design a target‑user prediction system
Company: Stripe
Role: Data Scientist
Category: Machine Learning
Difficulty: hard
Interview Round: Technical Screen
You must identify users likely to adopt Product P in the next 30 days to prioritize outreach. Data: user_profile (static attrs), user_events (timestamped events: page_view, search, add_to_cart, purchase, unsubscribe, etc.), marketing_contacts (timestamps and channel), and product_catalog. Training window: 2025‑03‑01 to 2025‑06‑30; prediction window: 2025‑07‑01 to 2025‑07‑31. Answer: (1) Precisely define the prediction target and labeling rule while preventing target leakage (consider contacts and post‑label features). (2) Propose features (behavioral recency/frequency, content affinity, embeddings) with an explicit time cutoff, and how you’d handle cold‑start users. (3) Choose a model (ranking vs classification) and justify with pros/cons given class imbalance and outreach budget constraints. (4) Specify offline metrics (PR‑AUC, top‑k recall, calibration/Brier) and how they map to online business outcomes. (5) Given a daily outreach budget B that limits you to contacting at most 50,000 users/day, formulate threshold selection to maximize expected incremental profit: write the objective that uses p(adopt|contact), incremental lift, contact cost, and capacity constraint; explain how you’d estimate incremental lift from observational data. (6) Show a time‑series cross‑validation scheme that respects user and temporal leakage. (7) Detail calibration and post‑processing (e.g., isotonic, Platt), fairness constraints across markets, and drift detection/retraining triggers (e.g., PSI thresholds). (8) Outline ablation and slice‑robustness checks you’d include in the presentation to pre‑empt Q&A.
Quick Answer: This question evaluates skills in temporal label definition and leakage prevention, time‑bounded feature engineering (recency/frequency and content affinity), uplift and incremental effect estimation, model selection under class imbalance and budgeted outreach, offline metric design, calibration, fairness constraints, and production validation and monitoring. Common in Machine Learning/Data Science interviews, it is asked to gauge both conceptual understanding of causal and validation issues and practical application skills for deploying a ranked outreach model under capacity constraints, including appropriate cross‑validation and thresholding to map predictive scores to business outcomes.