Top 60 Data Science Interview Questions & Answers (Python, ML, Stats) 2026

Lists are mutable. You can change them after creation. Tuples are immutable. Once created, they stay the same. The practical implication is that lists are for collections that change, tuples are for fixed data. Tuples are also faster and can be used as dictionary keys.

A compact way to create lists. Instead of writing a for loop, you write the expression inside square brackets. It is faster and more readable once you get used to it. But do not nest them three levels deep. That becomes unreadable and you will confuse yourself six months later.

First, you figure out why they are missing. The reason matters. Then you either drop them with dropna if the missing data is minimal, or fill them with fillna using mean, median, mode, or a more sophisticated imputation method. The choice depends on the data and the problem. There is no universal correct answer. That is the point of the question.

Apply works on both Series and DataFrames. Map works only on Series. Apply is more flexible. Map is simpler for element-wise substitutions. I use map for simple replacements and apply for anything that needs a custom function.

The merge function. Like a JOIN in SQL. You specify the key column and the type of merge. Inner, outer, left, right. Understanding these joins is essential. Most real-world data requires merging multiple sources.

An anonymous, one-line function. Useful for small operations you do not want to define a full function for. They are handy inside apply statements. They become unreadable if you try to do too much in one line. Keep them simple.

It splits your data into training and testing sets. The model learns on the training data. You evaluate it on the testing data. This prevents overfitting. If you train and test on the same data, your evaluation is meaningless. The model has already seen the answers.

Pandas for data manipulation. NumPy for numerical operations. Matplotlib and Seaborn for visualization. Scikit-learn for machine learning. Statsmodels for statistical modeling. Name them. Then explain briefly when you use each one.

pd.read_csv. It sounds trivial. But interviewers ask this to see if you have ever actually opened a CSV file in Python. Mention parameters like encoding and handling bad lines. That signals real experience.

A Series is one column. A DataFrame is multiple columns. A DataFrame is essentially a collection of Series. Simple distinction. But it matters for understanding how Pandas structures data.

drop_duplicates. You can specify a subset of columns to check for duplicates. You can keep the first or last occurrence. Real data is full of duplicates that are not exact duplicates. Cleaning them requires judgment.

iloc uses integer positions. loc uses labels. This confuses beginners constantly. iloc for row number. loc for row index name. Knowing the difference prevents silent bugs.

You generally should not. Vectorized operations are faster. But if you must, iterrows or itertuples. itertuples is faster. Mention that loops should be a last resort in Pandas. That shows you understand performance.

An isolated Python environment for a project. It keeps dependencies separate. Without it, different projects with conflicting library versions break each other. This is one of those things that sounds boring but will save you from nightmare debugging sessions.

They want to see if you can code something simple without googling. Return string equals string reversed. Keep it clean. Use slicing. Show you can write basic logic without overcomplicating it.

The probability of observing results at least as extreme as yours, assuming the null hypothesis is true. A low p-value suggests your results are unlikely to be due to chance. But p-value is not the probability that the null hypothesis is false. That misinterpretation is everywhere. Knowing the difference signals real understanding.

Correlation means two variables move together. Causation means one variable directly affects the other. Ice cream sales and drowning deaths are correlated. Both increase in summer. Ice cream does not cause drowning. Heat causes both. Confounding variables are the reason correlation is not causation.

The distribution of sample means approaches a normal distribution as sample size increases, regardless of the original distribution. This is why many statistical methods work. It is the foundation of inferential statistics. Know it intuitively, not just mathematically.

Type I is a false positive. Rejecting a true null hypothesis. Type II is a false negative. Failing to reject a false null hypothesis. The tradeoff between them depends on the context. In medical testing, a false negative might be worse than a false positive. In spam detection, a false positive, good email marked as spam, is worse than a false negative.

Standard deviation measures variability in your data. Standard error measures variability in your sample mean estimate. Standard error is always smaller. It decreases with sample size. Confusing them is a common beginner mistake.

A range that likely contains the true population parameter. A 95 percent confidence interval means that if you repeated the sampling process many times, 95 percent of the intervals would contain the true value. It is not the probability that the true value is in your specific interval. That distinction is subtle but important.

A symmetric, bell-shaped distribution defined by mean and standard deviation. Many natural phenomena approximate it. Many statistical tests assume it. But real data is often not normal. Knowing when the assumption holds is more important than knowing the formula.

Box plots. Z-scores. IQR method. But detection is easy. Deciding what to do with outliers is hard. Are they errors? Are they genuine extreme values? The context decides whether you remove, transform, or keep them.

Skewness measures asymmetry. Kurtosis measures tailedness. A positive skew means a long right tail. Income data is positively skewed. High kurtosis means heavy tails. More extreme values than a normal distribution. These properties affect which statistical methods are appropriate.

It updates the probability of a hypothesis based on new evidence. Prior probability plus new data equals posterior probability. It is the mathematical formalization of learning from experience. The formula looks intimidating. The concept is intuitive.

Bias is error from oversimplification. Variance is error from oversensitivity to training data. High bias underfits. High variance overfits. The tradeoff is finding the sweet spot. This is the single most important concept in machine learning. If you truly understand it, you understand model selection.

The model learns noise instead of signal. It performs great on training data and terribly on new data. Prevent it with cross-validation, regularization, simpler models, more data, and early stopping. Overfitting is the most common mistake in applied machine learning.

Splitting data into multiple folds. Training on some, validating on others. Rotating which fold is the validation set. K-fold is the standard. It gives a more reliable estimate of model performance than a single train-test split.

Supervised has labeled data. You know the answer you are trying to predict. Regression and classification. Unsupervised has unlabeled data. You are finding patterns without predefined answers. Clustering and dimensionality reduction.

It splits data based on features to create homogeneous groups. Each split tries to maximize information gain or minimize impurity. It is intuitive and interpretable. But single trees overfit easily. That is why ensembles like random forests exist.

An ensemble of decision trees. Each tree is trained on a random subset of data and features. The final prediction is the average or majority vote. It reduces overfitting. It handles non-linear relationships well. It is the workhorse algorithm for structured data.

It builds trees sequentially. Each new tree corrects the errors of the previous ones. XGBoost, LightGBM, CatBoost are implementations. They are powerful and often win competitions. They are also prone to overfitting if not tuned properly.

Bagging trains models in parallel. Each model is independent. Reducing variance is the goal. Boosting trains models sequentially. Each model learns from previous errors. Reducing bias is the goal. Random forest is bagging. XGBoost is boosting.

Accuracy, precision, recall, F1-score. But accuracy is misleading for imbalanced datasets. If 95 percent of samples are class A, a model that always predicts A has 95 percent accuracy and is useless. Use precision and recall. Use the confusion matrix. Understand the tradeoffs.

ROC plots true positive rate against false positive rate at different thresholds. AUC is the area under it. Higher AUC means better discrimination. AUC of 0.5 is random. AUC of 1.0 is perfect. It is a good overall metric for balanced classification.

Resampling. Oversample the minority class or undersample the majority. SMOTE is a popular oversampling technique. Use appropriate metrics like precision-recall instead of accuracy. Use algorithms that handle imbalance well. This problem is extremely common in real-world data.

Adding a penalty to model complexity. L1, Lasso, can reduce coefficients to zero, performing feature selection. L2, Ridge, shrinks coefficients but keeps all features. Regularization prevents overfitting. It is a fundamental technique.

K-means partitions data into K clusters. You specify K. It is fast and scales well. Hierarchical clustering builds a tree of clusters. You do not need to specify K. It is slower and does not scale to huge datasets. Both have their place.

Principal Component Analysis. Reduces dimensionality by finding directions of maximum variance. It transforms correlated features into uncorrelated components. Used for visualization, noise reduction, and speeding up other algorithms. The components are linear combinations of original features, which makes interpretation tricky.

Creating new features from existing ones to improve model performance. It is where domain expertise meets data science. It is often more impactful than algorithm choice. A good feature can boost performance more than switching from a basic model to an advanced one.

Discriminative models learn the boundary between classes. Generative models learn the distribution of each class. Naive Bayes is generative. Logistic regression is discriminative. Discriminative often performs better with enough data. Generative can work with less.

Filter methods. Statistical tests. Wrapper methods. Recursive feature elimination. Embedded methods. Lasso regularization. Domain knowledge. Correlation analysis. The best approach combines automated methods with human judgment.

Use a subquery with LIMIT and OFFSET, or use a window function like DENSE_RANK. This is a classic. It tests whether you can think beyond basic SELECT statements.

WHERE filters rows before aggregation. HAVING filters groups after aggregation. You cannot use aggregate functions in WHERE. You use HAVING for that. This distinction comes up constantly in real queries.

INNER JOIN returns matching rows from both tables. LEFT JOIN returns all rows from the left table and matching rows from the right. RIGHT JOIN is the reverse. FULL OUTER JOIN returns all rows from both. Knowing when to use each prevents data loss and incorrect results.

A query inside another query. Used in WHERE, FROM, or SELECT clauses. They can be powerful. They can also be slow if not written carefully. Sometimes a JOIN is better. Knowing the tradeoffs matters.

Check the execution plan. Add indexes. Avoid SELECT asterisk. Filter early with WHERE. Limit subqueries. Use appropriate JOIN types. Query optimization is a practical skill that separates people who have worked with real data from those who have only run tutorials.

Functions that perform calculations across a set of rows related to the current row. ROW_NUMBER, RANK, LAG, LEAD. They are incredibly useful for running totals, rankings, and comparisons without losing row-level detail.

A column or set of columns that uniquely identifies each row. It cannot contain NULL values. Every table should have one. It is fundamental to database design.

A column that creates a relationship between two tables. It references the primary key of another table. It enforces referential integrity. It prevents orphaned records.

Use IS NULL or IS NOT NULL. Not equals NULL because NULL is not a value, it is the absence of a value. COALESCE to replace NULLs with a default. NULL handling trips up even experienced people.

GROUP BY the columns that should be unique. Use HAVING COUNT asterisk greater than one. This is a practical data cleaning task that comes up constantly.

Use the framework. Problem, approach, challenges, results, lessons learned. Do not just describe what you did. Explain why you made the choices you made. The why matters more than the what.

Give a real failure. Not a humble brag. Explain what went wrong and what you learned. This tests honesty and the ability to learn from mistakes. Everyone makes errors. Not everyone can talk about them openly.

Consider business impact and urgency. Communicate timelines clearly. Do not overpromise. If everything is high priority, nothing is. Ask stakeholders to clarify the cost of delay for each request.

Name specific sources. A newsletter you read. A conference you attended. A paper you found interesting. Then say something about applying what you learn. Passive consumption is not impressive. Active experimentation is.

Pick something real. Explain why it was hard. Not just technically hard. Maybe the data was messy. Maybe the stakeholder kept changing requirements. Maybe the deadline was impossible. Show how you navigated the difficulty.

Start by quantifying the poorness. What exactly is wrong. Missing values, inconsistencies, outliers. Then prioritize fixes based on impact. Communicate limitations to stakeholders. A model built on bad data is worse than no model at all.

Be honest. Not "because it is the sexiest job of the 21st century." Say something real. The satisfaction of finding answers in data. The variety of problems. The constant learning. Something that sounds like a human being, not a LinkedIn headline.