Design a model to predict user retention based on interaction data.

Example answer

I would specify an actionable retention definition and scoring time. For a new-user example, use interactions in days 0–6 after signup to predict whether the account completes a meaningful activity in days 28–34. Define day boundaries, timezone, eligibility and the activity event. This is exact-window retention, not continuous use or permanent absence of churn.

Create one dated feature snapshot per eligible account, using permitted counts, recency, task completion and error events from the observation window. Deduplicate retries and distinguish missing telemetry from no activity. Wait until the label window is complete; an account observed only through day 20 is not a nonretained example. Keep future behavior and post-score interventions out of the features.

Start with a constant-probability or simple rule baseline and a regularized logistic model. Evaluate later signup cohorts, keeping related accounts together where needed. Fit preprocessing on training data, tune separately and reserve a final evaluation. Check calibration and ranking, precision/recall at the intended action capacity, and results by relevant cohort. Do not force a 50:50 test set and then interpret its precision as representative of actual prevalence.

Deploy a versioned batch score or endpoint with the same feature definitions, freshness checks, monitoring and a fallback when inputs are incomplete. Record score time and model version so matured outcomes can be linked correctly. A retention score predicts an outcome; it does not establish that encouraging a correlated behavior will cause retention. Evaluate any follow-up intervention separately, and validate retraining or threshold changes before rollout.