Data Scientist Interview Questions & Answers (with Model Answers)

Model Answer

I built a churn prediction model for a subscription product where retention was the key metric. Rather than chasing accuracy, I framed it around the business goal of prioritising retention outreach, so I optimised for precision at the top decile the team could actually action. I worked with marketing to A/B test the intervention, which lifted retention in the targeted group by several points. The lesson I emphasise is that the model only mattered because it changed a decision.

Model Answer

I start from the problem framing, the data available, and the interpretability and latency requirements rather than reaching for the fanciest model. I usually establish a simple baseline like logistic regression first, because it sets a bar and is easy to explain to stakeholders. I only add complexity if it earns its keep on a proper validation set. I also weigh maintainability, since a slightly better model that no one can debug is often the wrong choice.

Model Answer

I lead with the answer and the recommended action, then support it with the minimum evidence needed rather than walking through the maths. I translate metrics into business terms, for example expected revenue impact instead of AUC. I am explicit about uncertainty and assumptions so decisions are made with eyes open. I use a clear visual where it helps and invite questions to confirm understanding.

Model Answer

I version my code and data, fix random seeds, and keep a clear pipeline from raw data to result so anyone can rerun it. I validate assumptions, check for leakage between training and evaluation, and sanity-check results against known benchmarks. I document the methodology and its limitations honestly. Peer review of analysis is as important to me as code review.

Model Answer

Leadership believed a popular feature drove conversion, but my analysis showed the correlation vanished after controlling for user intent. I double-checked the result and pre-empted objections by testing alternative explanations. I presented it carefully, acknowledging the surprise and showing the evidence step by step. They ran a controlled experiment that confirmed my finding and reallocated resources accordingly.

A Type I error is a false positive, rejecting a true null hypothesis, controlled by the significance level alpha. A Type II error is a false negative, failing to reject a false null hypothesis, related to statistical power. There is a trade-off between them, and the right balance depends on the relative cost of each error in context.

Bias is error from overly simple assumptions that underfit, while variance is error from sensitivity to the training data that overfits. Increasing model complexity lowers bias but raises variance, and the goal is to minimise total error on unseen data. Techniques like regularisation, cross-validation, and more data help find the sweet spot.

Cross-validation splits the data into folds, training on some and validating on the held-out fold, then rotating so every observation is validated once. Averaging across folds gives a more reliable estimate of out-of-sample performance than a single split. It is especially valuable with limited data and for tuning hyperparameters without overfitting to one validation set.

Accuracy is misleading on imbalanced data, where predicting the majority class can score high while being useless. Precision measures how many predicted positives are correct, and recall measures how many actual positives were caught. You choose based on cost, for example high recall for disease screening and high precision for spam filtering.

Regularisation penalises model complexity to reduce overfitting by adding a term to the loss function. L1, or Lasso, penalises the absolute value of coefficients and can shrink some to zero, performing feature selection. L2, or Ridge, penalises squared coefficients and shrinks them smoothly without eliminating features, which handles correlated predictors well.

I was handed transaction data with inconsistent formats and large gaps. I profiled the data to quantify the missingness, then decided per column whether to impute, drop, or flag based on the mechanism behind it. I documented every cleaning decision so results were defensible. The cleaned dataset became the team's trusted source and my documentation saved others from redoing the work.

A model that looked strong offline degraded after deployment because the live feature distribution had drifted from training data. I set up monitoring on input distributions and the prediction quality to catch it. I retrained on recent data and added a scheduled retraining pipeline. Performance recovered and the monitoring caught the next drift before it hurt users.

Leadership needed a market-sizing estimate before a board meeting in two days. I scoped to a defensible back-of-envelope model using the most reliable available data rather than a perfect one. I was transparent about the confidence intervals and assumptions. The estimate informed the decision on time, and I noted what a deeper follow-up would refine.

A team was shipping features and declaring success from before-and-after numbers contaminated by seasonality. I showed how a controlled A/B test would isolate the real effect and ran one as a pilot. The test revealed a feature they thought was a win was actually flat. The team adopted experimentation as the default for major launches.