Data Scientist Interview Questions

While analysing retention cohorts, I discovered users who completed 3+ actions in their first session retained at 73% vs 31% for users who completed 1 action. This 'magic number 3' insight led to a complete onboarding redesign that increased 30-day retention by 22%. I presented the cohort analysis with confidence intervals to the CPO, which led to it becoming a Q3 company OKR.

L1 (Lasso) adds the absolute value of coefficients as a penalty, which drives some coefficients to exactly zero — producing sparse models and performing implicit feature selection. Use L1 when you suspect many features are irrelevant. L2 (Ridge) adds the squared value of coefficients, which shrinks all coefficients towards zero but rarely eliminates them entirely. Use L2 when all features are likely relevant and you want to reduce variance without feature elimination. ElasticNet combines both.

Signs of data leakage: (1) Suspiciously high cross-validation accuracy (>95% for real-world problems). (2) Feature importance dominated by a feature that wouldn't be available at prediction time. (3) Performance degrades sharply on out-of-time validation vs cross-validation. Prevention: use time-aware splits for time-series data, never include target-derived features, and audit your pipeline step by step — particularly preprocessing steps applied before train/test split.

I ask 3 questions: (1) What decision depends on this result? (2) What accuracy level is acceptable for that decision? (3) Can I give a 70% confidence answer tomorrow vs a 95% answer in 2 weeks? If they need directional guidance rather than final numbers, I can run exploratory analysis with clearly stated assumptions and confidence levels in 24 hours and present it as 'initial signal, not final answer.' I always communicate uncertainty explicitly — stakeholders respect honest limitations more than inflated precision.