Data Scientist Interview Questions & Answers
Data science interviews combine statistics, machine learning, programming, and business acumen. This comprehensive guide covers questions from foundational concepts to advanced ML techniques.
Statistics & Probability
Q1: Explain the difference between Type I and Type II errors.
Type I Error (False Positive): Rejecting the null hypothesis when it's actually true. Example: Concluding a drug is effective when it isn't. Probability is α (significance level, typically 0.05).
Type II Error (False Negative): Failing to reject the null hypothesis when it's actually false. Example: Concluding a drug isn't effective when it actually is. Probability is β. Power = 1 - β.
Trade-off: Reducing Type I error (lower α) increases Type II error, and vice versa. Choose based on which error is more costly in your context.
Q2: What is p-value and how do you interpret it?
The p-value is the probability of observing results as extreme as the data, assuming the null hypothesis is true.
Common misconception: P-value is NOT the probability that the null hypothesis is true.
Interpretation: If p < α (e.g., 0.05), we reject the null hypothesis. This means the observed result would be unlikely if the null were true.
Limitations: P-values don't measure effect size or practical significance. A tiny, meaningless difference can have a small p-value with large sample sizes.
Q3: Explain Bayes' Theorem with an example.
Bayes' Theorem: P(A|B) = P(B|A) × P(A) / P(B)
Example: A disease affects 1% of the population. A test has 95% sensitivity (true positive rate) and 90% specificity (true negative rate). If you test positive, what's the probability you have the disease?
P(Disease|Positive) = P(Positive|Disease) × P(Disease) / P(Positive) = (0.95 × 0.01) / (0.95 × 0.01 + 0.10 × 0.99) = 0.0095 / 0.1085 ≈ 8.8%
Despite the positive test, there's only an 8.8% chance of having the disease because the disease is rare (low base rate).
Q4: What is the Central Limit Theorem?
The CLT states that the sampling distribution of the sample mean approaches a normal distribution as sample size increases, regardless of the population's distribution (assuming finite variance).
Why it matters: Enables statistical inference even when we don't know the population distribution. With n ≥ 30, the sample mean is approximately normally distributed.
Applications: Confidence intervals, hypothesis tests, quality control.
Machine Learning Fundamentals
Q5: Explain the bias-variance trade-off.
Bias: Error from oversimplified assumptions. High bias = underfitting (model too simple to capture patterns).
Variance: Error from sensitivity to training data fluctuations. High variance = overfitting (model too complex, memorizes training data).
Trade-off: Decreasing bias often increases variance and vice versa. Goal is to find the sweet spot that minimizes total error.
Solutions for high bias: More features, more complex model, less regularization.
Solutions for high variance: More data, fewer features, regularization, ensemble methods.
Q6: How do you handle imbalanced datasets?
Data-level approaches:
- Oversampling minority class (SMOTE, random oversampling)
- Undersampling majority class
- Generating synthetic samples
Algorithm-level approaches:
- Class weights (penalize misclassifying minority class more)
- Threshold adjustment
- Anomaly detection approach
Evaluation: Don't use accuracy. Use precision, recall, F1-score, AUC-ROC, or precision-recall curves.
Q7: Explain regularization and compare L1 vs L2.
Regularization adds a penalty term to the loss function to prevent overfitting by constraining model complexity.
L1 (Lasso): Adds |weights| to loss. Produces sparse models (some weights become exactly 0). Good for feature selection.
L2 (Ridge): Adds weights² to loss. Shrinks weights toward 0 but rarely exactly 0. Better when most features are relevant.
Elastic Net: Combines L1 and L2 for benefits of both.
Q8: What is cross-validation and why use it?
Cross-validation is a technique to assess model performance on unseen data by partitioning data into training and validation sets multiple times.
K-Fold CV: Split data into k folds, train on k-1, validate on 1, rotate through all folds. Average performance across folds.
Why use it: More reliable performance estimate than single train-test split, uses all data for both training and validation, essential for hyperparameter tuning.
Stratified CV: Maintains class proportions in each fold (important for imbalanced data).
Advanced Machine Learning
Q9: Explain gradient boosting.
Gradient boosting builds an ensemble of weak learners (typically decision trees) sequentially, where each new learner corrects the errors of the combined ensemble so far.
Process:
- Initialize with a simple prediction (e.g., mean)
- Calculate residuals (errors)
- Fit a new tree to predict residuals
- Add tree to ensemble (scaled by learning rate)
- Repeat
Key parameters: Number of trees, learning rate (smaller = more robust but needs more trees), tree depth (shallower = less overfitting).
Popular implementations: XGBoost, LightGBM, CatBoost.
Q10: How do neural networks learn?
Neural networks learn through backpropagation and gradient descent:
- Forward pass: Input flows through network, producing output
- Loss calculation: Compare prediction to actual value
- Backward pass: Compute gradients of loss with respect to each weight using chain rule
- Weight update: Adjust weights in direction that reduces loss
Key concepts: Learning rate (step size), activation functions (introduce non-linearity), loss functions (define what to optimize).
Q11: Explain the attention mechanism in transformers.
Attention allows models to focus on relevant parts of the input when producing each output element.
Self-attention: Each position attends to all positions in the same sequence. Computes Query, Key, Value vectors from input, then:
Attention(Q, K, V) = softmax(QK^T / √d_k) × V
Why it works: Captures long-range dependencies regardless of distance (unlike RNNs). Parallelizable (unlike sequential models).
Multi-head attention: Multiple attention "heads" learn different types of relationships, outputs are concatenated.
Q12: What is the difference between generative and discriminative models?
Discriminative models learn the boundary between classes (P(Y|X)). Examples: Logistic regression, SVM, neural networks for classification.
Generative models learn the distribution of each class (P(X|Y) and P(Y)). Examples: Naive Bayes, Gaussian Mixture Models, GANs.
Trade-offs: Discriminative models often perform better for classification when there's enough data. Generative models can generate new samples, handle missing data, and work with less labeled data.
Practical & Applied Questions
Q13: How would you approach a new data science problem?
Framework:
Understand the problem: What's the business goal? How will the model be used? What's the baseline?
Data exploration: Quality, distributions, missing values, relationships
Feature engineering: Domain knowledge, transformations, interactions
Model selection: Start simple, iterate to complexity
Evaluation: Appropriate metrics, cross-validation, holdout set
Deployment: Infrastructure, monitoring, maintenance plan
Iteration: Feedback loop, continuous improvement
Q14: How do you handle missing data?
First: Understand why data is missing (MCAR, MAR, MNAR)
Options:
- Deletion: Remove rows/columns (if missing completely at random and small %)
- Simple imputation: Mean, median, mode (can reduce variance)
- Advanced imputation: KNN imputation, MICE (multiple imputation), model-based
- Use as signal: Create "missing" indicator feature
- Models that handle missing: Some tree-based models handle missing values natively
Q15: Your model performs well in testing but poorly in production. Why?
Common causes:
Data drift: Production data differs from training data distribution
Feature drift: Features calculated differently or unavailable
Label leakage: Training data had information not available at prediction time
Feedback loops: Model predictions influence future training data
Evaluation mistake: Test data wasn't truly held out (time-based leakage)
Solution: Monitor prediction distributions, feature distributions, and model performance continuously. Have clear rollback procedures.
Q16: How would you explain a machine learning model to a non-technical stakeholder?
Principles:
- Lead with business impact, not technical details
- Use analogies they can relate to
- Visualize when possible
- Be honest about limitations and uncertainty
Example for gradient boosting: "The model is like a team of experts where each one corrects the mistakes of the previous experts. Instead of one person making all decisions, we combine many focused insights to make better predictions overall."
Coding & SQL Questions
Q17: How would you find duplicate records in a large dataset?
SQL approach:
SELECT column1, column2, COUNT(*) as count
FROM table
GROUP BY column1, column2
HAVING COUNT(*) > 1
Python approach:
duplicates = df[df.duplicated(subset=['column1', 'column2'], keep=False)]
# or
df.groupby(['column1', 'column2']).filter(lambda x: len(x) > 1)
Q18: Explain your approach to feature engineering.
Categories of feature engineering:
Numerical: Scaling, binning, log transforms, polynomial features
Categorical: One-hot encoding, label encoding, target encoding, frequency encoding
Temporal: Extracting date parts, cyclic encoding, lag features, rolling statistics
Text: TF-IDF, word embeddings, n-grams
Domain-specific: Ratio features, interaction terms, aggregations
Key principle: Feature engineering is often more impactful than model selection. Invest time understanding the domain.
This guide covers essential data science interview topics. Remember to explain your reasoning, acknowledge limitations, and connect technical concepts to business outcomes. Practice articulating complex ideas simply—it's a skill that distinguishes great data scientists.