Having the schema handed to you sounds like a gift but it kind of paralyzed me at first because I kept second-guessing which features were actually signal vs noise.
Start by clarifying the schema and defining churn precisely (e.g., no ride in X weeks), then walk through a structured ML pipeline: data exploration, feature engineering, model selection, validation, and deployment. Emphasize how you'd handle class imbalance, temporal validation, and business impact.
Pro tip: Frame churn as a time-dependent problem and use time-based splits rather than random splits to avoid leakage; also discuss how you'd translate model outputs into actionable retention campaigns.
Ask about available tables (rides, user profile, app events) and define churn operationally, e.g., no ride in the next 4 weeks. Confirm the prediction window and observation window.
Perform EDA to understand distributions and missingness. Create features like ride frequency trends, recency, monetary value, app engagement, and tenure, ensuring they are computed only from data before the prediction point.
Choose interpretable baselines (logistic regression) and tree-based models (XGBoost/LightGBM) to handle non-linearity. Address class imbalance with techniques like class weights or resampling, and tune hyperparameters.
Use temporal validation (e.g., train on earlier weeks, validate on later weeks) to mimic production. Evaluate with precision-recall AUC, lift at top decile, and calibration, not just accuracy.
Discuss how to productionize the model (batch scoring, API), set up monitoring for drift and performance, and integrate with retention systems to trigger interventions.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.