Data Scientist Interview Questions & Answers

Sample Answer

Bias measures how far predictions are from true values on average—high bias means the model is too simple and underfits. Variance measures how much predictions change with different training data—high variance means the model overfits to noise. The tradeoff exists because reducing one often increases the other. Simple models have high bias, low variance; complex models have low bias, high variance. The goal is finding the sweet spot that minimizes total error. I manage this through cross-validation, regularization, and ensemble methods that combine multiple models.

Sample Answer

First, I analyze the missing data pattern—is it random, systematic, or related to other variables? For random missing data, I might impute with mean/median for numerical columns or mode for categorical. For more sophisticated imputation, I use techniques like KNN imputation or multiple imputation. If data is missing not at random (MNAR), imputation can introduce bias, so I might create a missing indicator feature or use models that handle missing values natively like XGBoost. For significant missingness (>30%), I consider dropping the column. I always validate imputation doesn't distort distributions.

Sample Answer

L1 (Lasso) adds the absolute value of coefficients as a penalty, encouraging sparsity by driving some coefficients to exactly zero—useful for feature selection. L2 (Ridge) adds squared coefficients, shrinking all weights toward zero but rarely eliminating features entirely—better when all features are potentially relevant. L1 produces more interpretable models; L2 handles correlated features better. Elastic Net combines both. I choose L1 when I suspect many irrelevant features, L2 when features are correlated, and Elastic Net when unsure.

Sample Answer

I built a customer churn prediction model for a SaaS company. Starting with exploratory analysis of 50K customers and 200 features, I identified key predictors: login frequency decline, support tickets, and contract end dates. I engineered features like rolling averages and recency scores. After testing logistic regression, random forest, and XGBoost, XGBoost performed best with 0.85 AUC. I optimized the threshold for business constraints—maximizing recall for high-value customers. The model helped the retention team prioritize outreach, reducing churn by 15% and saving $500K annually.

Sample Answer

I'd use a subquery approach: SELECT MAX(salary) FROM employees WHERE salary < (SELECT MAX(salary) FROM employees). Alternatively, using window functions: SELECT salary FROM (SELECT salary, DENSE_RANK() OVER (ORDER BY salary DESC) as rank FROM employees) ranked WHERE rank = 2. The window function approach handles ties better and is more extensible for nth highest. I'd add DISTINCT if there might be duplicate salaries. For production code, I'd also handle edge cases like tables with fewer than two distinct salaries.

Sample Answer

The choice depends on the business problem. Accuracy is misleading for imbalanced classes. Precision matters when false positives are costly (spam detection); recall matters when false negatives are costly (disease detection). F1 score balances both. AUC-ROC shows performance across thresholds, useful for ranking. I also examine confusion matrices and precision-recall curves. For a fraud detection model I built, we prioritized precision at high recall thresholds because investigating false positives was acceptable but missing fraud was costly. Business context always drives metric selection.

Sample Answer

Random Forest is an ensemble of decision trees that combines their predictions through voting (classification) or averaging (regression). Each tree is trained on a bootstrap sample of the data and considers only a random subset of features at each split, introducing diversity. This reduces overfitting compared to single trees and provides feature importance through measuring how much each feature reduces impurity. I use it when I need good performance without extensive tuning, interpretable feature importance, and robustness to outliers. Limitations include difficulty with extrapolation and larger model size.

Sample Answer

I focus on what the model does and its business impact rather than technical details. For a churn model, I'd explain: "This model identifies customers likely to cancel in the next 30 days based on their behavior patterns. It examines factors like how often they log in, support tickets, and payment history. It's about 85% accurate, meaning for every 100 customers it flags, 85 will actually churn if we don't intervene." I use visualizations, avoid jargon, and always connect to decisions: "This lets us prioritize which customers to call, saving 20 hours weekly while reducing churn."

Sample Answer

As dimensions increase, data becomes sparse—points that seemed close in low dimensions become far apart. This causes problems: distance metrics lose meaning, models need exponentially more data to maintain coverage, and overfitting becomes likely. For example, 10 points densely cover a 1D line but are extremely sparse in 100 dimensions. I address this through feature selection to remove irrelevant features, dimensionality reduction like PCA or t-SNE for visualization, and regularization. I also use tree-based models that are more robust to high dimensions than distance-based methods like KNN.

Sample Answer

Several approaches depending on the severity and problem. Resampling: oversampling minority class (SMOTE) or undersampling majority class. Cost-sensitive learning: assigning higher misclassification costs to minority class. Algorithm selection: tree-based methods and anomaly detection handle imbalance better. Evaluation: use precision-recall curves and F1 rather than accuracy. Threshold tuning: adjust classification threshold based on business costs. For a fraud detection project with 0.1% fraud rate, I combined SMOTE with XGBoost and tuned the threshold to achieve 80% recall while maintaining acceptable precision.

Sample Answer

A/B testing compares two variants by randomly assigning users and measuring outcome differences. Common pitfalls include: stopping tests too early (peeking problem), testing too many variants without correction for multiple comparisons, not accounting for novelty effects, contamination between groups, and using wrong statistical tests. I ensure adequate sample size calculation upfront, use sequential testing methods if early stopping is needed, apply Bonferroni correction for multiple tests, and wait sufficient time for effects to stabilize. I also segment results to check for different effects across user groups.

Sample Answer

Cross-validation assesses model performance by training and testing on different data subsets. K-fold CV splits data into K parts, trains on K-1, tests on the remaining fold, and rotates through all combinations. It's important because it gives a more reliable performance estimate than a single train-test split, helps detect overfitting, and makes efficient use of limited data. I typically use 5 or 10 folds. For time series, I use time-based splits to prevent data leakage. Stratified CV maintains class proportions. Cross-validation guides hyperparameter tuning without contaminating the test set.

Sample Answer

I follow a structured process: First, understand the business problem deeply—what decision will this enable? Then examine available data: sources, quality, and volume. EDA reveals patterns, distributions, and data issues. I establish baseline models and success metrics aligned with business goals. Feature engineering based on domain knowledge often matters most. I iterate through models, starting simple and adding complexity only if needed. I validate thoroughly and consider deployment requirements early. Finally, I communicate results clearly with actionable recommendations. This process helped me avoid building technically impressive but useless models.

Sample Answer

Gradient descent is an optimization algorithm that iteratively adjusts parameters to minimize a loss function. It calculates the gradient (direction of steepest increase) and takes steps in the opposite direction. Learning rate controls step size—too large causes overshooting, too small is slow. Variants include batch (uses all data, stable but slow), stochastic (one sample, fast but noisy), and mini-batch (balanced). Advanced optimizers like Adam adapt learning rates per parameter. I monitor loss curves to detect issues: oscillation suggests high learning rate, slow convergence suggests low rate or poor initialization.

Sample Answer

Supervised learning trains on labeled data to predict outcomes—classification for categories, regression for continuous values. Examples: spam detection, price prediction. Unsupervised learning finds patterns in unlabeled data—clustering groups similar items, dimensionality reduction compresses features. Examples: customer segmentation, anomaly detection. I choose supervised when I have labeled outcomes and want predictions; unsupervised for exploration and pattern discovery. Often I combine them: use clustering to create features for supervised models, or use classification to label data for further analysis.

Sample Answer

Detection methods include statistical approaches (Z-score, IQR), visualization (box plots, scatter plots), and model-based methods (isolation forest, DBSCAN). Handling depends on context: outliers might be errors to remove, valid extreme values to keep, or important anomalies to investigate. For errors, I impute or remove. For valid extremes, I might transform data (log), use robust statistics (median instead of mean), or use models robust to outliers (tree-based). For fraud detection, outliers are exactly what I'm looking for. I never blindly remove outliers without understanding their source.