Data Scientist Interview Questions & Answers (2026)

Sample Answer

Bias is the error from overly simplistic assumptions — a linear model on non-linear data has high bias and underfits. Variance is the error from sensitivity to noise — a deep decision tree memorizes training data and overfits. The total error is bias² + variance + irreducible noise. The sweet spot is a model complex enough to capture the signal but regularized enough to generalize. Techniques like cross-validation help find this balance empirically.

Sample Answer

I'd start with EDA on historical churned vs. retained users to identify discriminating features: usage frequency, support ticket volume, billing issues, feature adoption. I'd engineer features, handle class imbalance with SMOTE or class weights, and train a gradient boosting model (XGBoost). I'd evaluate with AUC-ROC and precision-recall curves, not accuracy — churn datasets are usually imbalanced. I'd deploy with a threshold tuned to business cost of false positives vs. false negatives.

Sample Answer

Regularization adds a penalty term to the loss function to discourage complex models and prevent overfitting. L1 (Lasso) adds the absolute value of coefficients, driving irrelevant features to exactly zero — useful for feature selection. L2 (Ridge) adds squared coefficients, shrinking all weights toward zero — better when all features matter. ElasticNet combines both. I use regularization whenever I have many features relative to training samples.

Sample Answer

I go beyond accuracy. I use a confusion matrix to understand false positive and false negative rates separately. I plot the ROC-AUC curve for overall discrimination ability and the precision-recall curve when classes are imbalanced. I choose the operating threshold based on the business cost of each error type. For multi-class problems I report macro and micro F1 scores.

Sample Answer

A random forest is an ensemble of decision trees trained on bootstrap samples of the data, each tree seeing a random subset of features at each split. Predictions are aggregated by majority vote (classification) or averaging (regression). This combination of bagging and feature randomness reduces variance without increasing bias. Random forests are robust to overfitting and handle missing data well, though they're less interpretable than single trees.

Sample Answer

Parametric tests assume the data follows a specific distribution — usually normal — and use parameters like mean and variance. Examples: t-test, ANOVA. Non-parametric tests make no distributional assumptions and rank-order data instead. Examples: Mann-Whitney U, Kruskal-Wallis. I default to parametric tests when sample sizes are large (CLT applies), and use non-parametric for small samples or clearly non-normal distributions.

Sample Answer

Red flags include suspiciously high training accuracy, features with very high feature importance that seem too predictive, and a large gap between cross-validation and holdout performance. I audit features for temporal leakage — ensuring no future information bleeds into training. For time-series models, I use strict time-ordered splits, not random cross-validation. I also check if any ID or index columns were accidentally included.

Sample Answer

PCA projects data onto orthogonal components that maximize variance, good for visualization and removing multicollinearity. t-SNE and UMAP are non-linear methods better for visualizing cluster structure in high-dimensional data. Feature selection methods (RFE, Lasso) are preferable over extraction when interpretability matters. I use PCA for preprocessing continuous features before feeding into linear models, and UMAP for exploratory clustering analysis.

Sample Answer

First, I check whether the imbalance reflects reality — don't over-correct. My options: oversample the minority class (SMOTE), undersample the majority, use class weights in the loss function, or use anomaly detection framing instead of classification. I evaluate with precision-recall AUC, not ROC-AUC, since ROC-AUC is optimistic on imbalanced datasets. I never generate synthetic samples in the test set.

Sample Answer

I translate model output into a business action and outcome. Instead of 'the model has 87% AUC,' I say 'by flagging the top 10% riskiest customers, we can proactively retain 60% of accounts at risk of churning in the next 90 days.' I show the lift chart or cumulative gain chart to visualize model value over random selection. I'm upfront about limitations and confidence intervals.