Data Science Interview Questions & Answers (2026)
These interviews test your ability to translate data into actionable insights, evaluate statistical reasoning, and implement scalable models. Success comes from mastering core concepts, clearly explaining methodology, and demonstrating impact through real‑world examples. Focus on problem‑solving steps, trade‑offs, and how you communicate results to non‑technical stakeholders.
20 questions · updated Aug 29, 2026
Quick facts
| Typical rounds | Screening, technical coding, case study, and system design |
| Core skills evaluated | Statistics, machine learning, data wrangling, and storytelling |
| Preferred tools | Python, SQL, pandas, scikit‑learn, and cloud notebooks |
Questions
Beginner
Explain the bias‑variance trade‑off in model selection.
Bias measures error from erroneous assumptions in the learning algorithm, while variance measures error from sensitivity to small fluctuations in the training set. High bias leads to underfitting, missing patterns; high variance leads to overfitting, capturing noise. An optimal model balances both, achieving low total error. Interviewers expect you to illustrate with a simple example, such as a linear model (high bias) versus a deep tree (high variance), and discuss techniques like cross‑validation to find the sweet spot.
How would you handle a dataset with 30% missing values?
First, assess the missingness mechanism—MCAR, MAR, or MNAR—to choose an appropriate strategy. Simple imputation (mean/median) works for MCAR, while more sophisticated methods like K‑Nearest Neighbors or model‑based imputation handle MAR. If missingness is informative (MNAR), consider adding a missing indicator feature. Always validate the impact on model performance using a hold‑out set, and document the rationale for reproducibility.
What is the difference between L1 and L2 regularization?
L1 regularization adds the absolute value of coefficients to the loss function, encouraging sparsity by driving some weights to zero, which performs feature selection. L2 adds the squared magnitude, penalizing large weights but retaining all features, leading to smoother solutions. Interviewers look for you to explain the geometric intuition, impact on bias‑variance, and typical use cases—L1 for high‑dimensional data, L2 for multicollinearity mitigation.
Describe how you would evaluate a classification model beyond accuracy.
Accuracy can be misleading with imbalanced classes. Complement it with precision (positive predictive value) and recall (sensitivity) to capture false positive and false negative rates. The F1 score balances both, while the ROC‑AUC measures discrimination across thresholds. For business impact, discuss cost‑sensitive metrics or confusion matrix analysis. Interviewers expect you to choose metrics aligned with the problem’s risk profile.
What is a p‑value and how do you interpret it?
A p‑value quantifies the probability of observing data as extreme as the sample, assuming the null hypothesis is true. A small p‑value (typically <0.05) suggests evidence against the null, prompting rejection. However, it does not measure effect size or practical significance. Interviewers want you to stress that statistical significance does not imply business relevance and to mention confidence intervals as a complementary tool.
Intermediate
Explain the concept of a confusion matrix.
A confusion matrix is a 2×2 (or NxN for multiclass) table that records true positives, false positives, true negatives, and false negatives. It provides the raw counts needed to compute metrics like accuracy, precision, recall, and specificity. By visualizing where errors occur, you can diagnose model weaknesses, such as a bias toward a particular class, and guide threshold tuning or data rebalancing strategies.
How do you choose between a decision tree and a random forest?
Decision trees are interpretable and fast but prone to overfitting. Random forests aggregate many trees, reducing variance and improving generalization at the cost of interpretability and higher computational load. Choose a tree when model transparency is critical or data is small; opt for a random forest when accuracy outweighs interpretability and you have sufficient resources. Discuss OOB error estimation and feature importance as advantages of forests.
What is cross‑validation and why is it important?
Cross‑validation partitions data into k folds, training on k‑1 folds and validating on the remaining one, rotating through all folds. It provides an unbiased estimate of model performance on unseen data, mitigates overfitting, and helps tune hyperparameters. Interviewers expect you to mention common variants (k‑fold, stratified, leave‑one‑out) and trade‑offs between bias, variance, and computational cost.
Explain the difference between bagging and boosting.
Bagging builds multiple independent models on bootstrapped samples and aggregates predictions (e.g., random forest), reducing variance. Boosting builds models sequentially, each focusing on errors of the previous one (e.g., XGBoost), reducing bias and often achieving higher accuracy. Interviewers look for you to discuss how bagging is parallelizable, while boosting is sequential and more prone to overfitting if not regularized.
How would you detect and handle multicollinearity?
Compute correlation matrix or Variance Inflation Factor (VIF); VIF >10 signals problematic multicollinearity. To address it, drop redundant features, combine them via PCA, or apply regularization (L2). Explain that multicollinearity inflates coefficient variance, making interpretation unstable, but does not necessarily degrade predictive performance. Interviewers appreciate a systematic diagnostic‑remediation workflow.
Describe the steps you would take to build a recommendation system.
Start by defining the business goal (e.g., click‑through rate). Gather interaction data and preprocess it (filter noise, handle sparsity). Choose a modeling approach: collaborative filtering (user‑item matrix factorization), content‑based, or hybrid. Train using implicit feedback, evaluate with metrics like MAP or NDCG, and iterate with hyperparameter tuning. Finally, consider scalability (e.g., Spark) and online A/B testing for deployment.
What is the Central Limit Theorem and why does it matter in data science?
The Central Limit Theorem states that the sampling distribution of the mean of independent, identically distributed variables approaches a normal distribution as the sample size grows, regardless of the original distribution. This justifies using confidence intervals and hypothesis tests on sample means, even when data are non‑normal. Interviewers expect you to link it to practical tasks like estimating population parameters and constructing prediction intervals.
Advanced
Explain how gradient descent works for linear regression.
Gradient descent iteratively updates model coefficients by moving opposite to the gradient of the loss function (usually MSE). At each step, compute predictions, calculate residuals, and adjust weights proportional to the learning rate times the gradient. Convergence depends on learning rate choice and feature scaling. Interviewers look for you to discuss batch vs. stochastic variants, convergence criteria, and potential pitfalls like local minima (though convex loss avoids them).
What is regularization and how does it prevent overfitting?
Regularization adds a penalty term to the loss function that discourages large coefficients, effectively shrinking them toward zero. L1 (Lasso) encourages sparsity, while L2 (Ridge) distributes shrinkage across all features. By limiting model complexity, regularization reduces variance, improving generalization on unseen data. Interviewers expect you to discuss hyperparameter tuning (λ) via cross‑validation and the bias‑variance trade‑off introduced by regularization.
How do you evaluate a time‑series forecasting model?
Split data chronologically into training, validation, and test sets to respect temporal order. Use metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). Additionally, assess residual autocorrelation with Ljung‑Box test and plot forecast vs. actual. Explain the importance of avoiding look‑ahead bias and possibly employing rolling‑origin evaluation for robustness.
What is the difference between A/B testing and multivariate testing?
A/B testing compares two variants (control vs. treatment) to isolate the effect of a single change. Multivariate testing evaluates multiple factors simultaneously, testing all possible combinations to understand interaction effects. Multivariate tests require larger sample sizes due to combinatorial explosion. Interviewers want you to discuss when each is appropriate, statistical power considerations, and how to interpret interaction terms.
Explain the concept of embeddings and when you would use them.
Embeddings map high‑dimensional categorical or textual data into dense, low‑dimensional vectors that capture semantic similarity. They are learned via neural networks (e.g., word2vec, item2vec) or matrix factorization. Use embeddings when you have sparse inputs like words, users, or items, and need to preserve relational structure for downstream models. Interviewers expect you to mention training objectives (e.g., skip‑gram) and benefits like reduced dimensionality and improved generalization.
How would you approach feature engineering for a churn prediction model?
Start by defining churn and labeling historical data. Engineer usage frequency, recency, monetary value, and interaction patterns (e.g., session length). Create temporal aggregates (rolling windows) and derive ratios (e.g., active days/total days). Incorporate demographic and account metadata, then apply domain‑specific transformations like tenure bins. Validate features using importance scores from tree models and ensure they are not leakage‑prone. Explain that iterative experimentation and business insight drive effective feature sets.
What is the purpose of a ROC curve and how do you interpret it?
A ROC curve plots the true positive rate against the false positive rate at various classification thresholds, illustrating the trade‑off between sensitivity and specificity. The area under the curve (AUC) summarizes overall discriminative ability; 0.5 indicates random guessing, 1.0 perfect separation. Interviewers look for you to discuss how to select an operating point based on business costs and how ROC is insensitive to class imbalance compared to precision‑recall curves.
Describe how you would use Bayesian inference in a data science project.
Begin with a prior distribution reflecting existing knowledge about a parameter. Collect data and define a likelihood function. Apply Bayes' theorem to obtain the posterior, which updates beliefs incorporating evidence. Use the posterior for prediction, credible intervals, or decision‑making under uncertainty. Interviewers expect you to mention tools like PyMC3 or Stan, and scenarios such as A/B testing where prior information improves sample efficiency.
Common mistakes
- Skipping data cleaning and assuming raw data is ready for modeling
- Choosing metrics that don’t align with business objectives
- Over‑explaining algorithms without linking to the problem context
- Neglecting to discuss model interpretability and deployment considerations
Study plan
- Review core statistics and probability concepts; solve practice problems daily
- Master Python data‑stack (pandas, numpy, scikit‑learn) with hands‑on notebooks
- Practice coding ML algorithms and SQL queries on real datasets
- Simulate end‑to‑end case studies: data cleaning → feature engineering → modeling → evaluation
- Conduct mock interviews focusing on explaining reasoning and trade‑offs
FAQ
How much math should I know for a data science interview?
You need a solid grasp of linear algebra, calculus basics, probability, and statistics. Expect questions on matrix operations, gradient descent, hypothesis testing, and distribution properties. Depth varies by role; research the specific job description to prioritize topics.
What programming language is most preferred by interviewers?
Python dominates due to its rich ecosystem, but many firms also accept R or Scala. Be prepared to write clean, efficient code in Python, especially using pandas and scikit‑learn, and to explain any alternative language choices you make.
Should I bring a portfolio of projects to the interview?
Yes. A concise portfolio demonstrates practical experience. Highlight the problem statement, data sources, methodology, results, and business impact. Be ready to discuss challenges, trade‑offs, and how you iterated on the solution.
How important is model interpretability in interviews?
Very important for most business‑focused roles. Interviewers often ask you to explain a model to non‑technical stakeholders. Prepare to discuss feature importance, SHAP values, or simple surrogate models, and why interpretability matters for trust and compliance.
What are common pitfalls when answering case‑study questions?
Common pitfalls include jumping to a solution without structuring the problem, ignoring data quality, and failing to quantify impact. Use a clear framework: define objectives, outline data requirements, propose an approach, discuss evaluation, and consider deployment constraints.
Related
Ready for your next interview?
Download MiPrep AI. Load your resume and the job description. Show up ready.
Free tier · No credit card · macOS 14+ · Windows 10+
Free tier · No credit card · Runs on your Mac or Windows machine