Home/ Data Science Course/ Top 100 Interview Questions
100 Real Questions 5 Core Areas Freshers & Experienced

Top 100 Data Science Interview Questions by Experts

The questions we actually use in mock interview rounds at Prime Point, organised by topic, with straight answers, not textbook definitions.

Python 20 Qs
Machine Learning 20 Qs
Deep Learning & NLP 20 Qs
Statistics & Probability 20 Qs
Model Evaluation & Deployment 20 Qs

Want guided prep?

Our Data Science program includes 3 live mock interview rounds covering exactly these topics.

Book Free Demo →
Python for Data ScienceQ1–Q20
1What's the difference between a list and a tuple in Python?
Lists are mutable and can be changed after creation; tuples are immutable, which makes them faster and hashable, so they can be used as dictionary keys.
2What are Python decorators?
Functions that wrap another function to extend its behavior without modifying its actual code, commonly used for logging, timing, or access control.
3Explain generators and why they're memory-efficient.
Generators yield values one at a time using yield, so they don't hold the entire sequence in memory the way a list would.
4What is the Global Interpreter Lock (GIL)?
A mutex in CPython that allows only one thread to execute Python bytecode at a time, which limits true parallelism for CPU-bound multi-threaded code.
5Difference between multiprocessing and multithreading in Python?
Multiprocessing runs separate processes with their own memory space, bypassing the GIL for true parallelism; multithreading shares memory but is limited by the GIL for CPU-bound tasks.
6What is NumPy broadcasting?
It lets NumPy perform operations on arrays of different shapes by automatically expanding the smaller array without physically copying data, saving memory and time.
7Explain the difference between .values and .to_numpy() in Pandas.
Both extract the underlying NumPy array from a DataFrame or Series; .to_numpy() is the newer, more explicit method and is the recommended one going forward.
8What is the purpose of __init__ in a Python class?
It's the constructor method that runs automatically when a new object is instantiated, used to initialize the object's instance attributes.
9How would you handle a memory error while processing a large dataset in Python?
Process data in chunks, use more memory-efficient dtypes, switch to a library like Dask for out-of-core computation, or downsample if the analysis allows it.
10What's the difference between a regular function and a lambda function?
A lambda is a small anonymous function limited to a single expression; a regular function defined with def can contain multiple statements and is named.
11Explain *args and **kwargs.
*args lets a function accept a variable number of positional arguments as a tuple; **kwargs lets it accept variable keyword arguments as a dictionary.
12What's the difference between mutable and immutable objects in Python?
Mutable objects, like lists and dictionaries, can be changed after creation; immutable objects, like strings, tuples, and integers, cannot.
13How do you profile and speed up slow Python code?
Use profiling tools like cProfile to find bottlenecks, then optimize with vectorization, better algorithms, caching, or compiled extensions like Cython or Numba.
14What's the difference between Pandas' merge() and concat()?
merge() joins DataFrames based on common keys or columns like a SQL join; concat() simply stacks DataFrames along an axis without matching on any keys.
15Explain how Python's exception hierarchy works.
All exceptions inherit from BaseException, with common ones like ValueError and TypeError inheriting from Exception, letting you catch broad or specific error types.
16What is JSON normalization, and when would you use it in Pandas?
json_normalize() flattens nested JSON structures into a flat table, useful when working with API responses that contain nested fields.
17What's the difference between a Python module and a package?
A module is a single .py file; a package is a directory of modules containing an __init__.py file that groups related functionality together.
18How do you handle categorical variables in Python for modeling?
Common techniques include one-hot encoding, label encoding, or target encoding, chosen based on cardinality and whether the model requires numeric input.
19What's the difference between .apply() and .map() in Pandas?
.map() works element-wise on a Series only; .apply() works on both Series and DataFrames and can operate row-wise, column-wise, or element-wise.
20Explain what a context manager (the "with" statement) does.
It automatically handles setup and teardown, like opening and closing a file, ensuring resources are released properly even if an error occurs mid-way.
Machine LearningQ21–Q40
21What's the difference between supervised and unsupervised learning?
Supervised learning trains on labeled data to predict a known outcome; unsupervised learning finds patterns or groupings in data without labeled outcomes.
22Explain the bias-variance tradeoff.
High bias means a model is too simple and underfits; high variance means it's too complex and overfits. The goal is balancing both for the best generalization.
23What is overfitting, and how do you prevent it?
Overfitting is when a model learns noise in the training data instead of the underlying pattern; it's prevented with regularization, cross-validation, more data, or simpler models.
24What is cross-validation?
A technique that splits data into multiple folds, training and testing the model on different combinations, to get a more reliable estimate of its performance.
25Explain the difference between L1 and L2 regularization.
L1 (Lasso) adds the absolute value of coefficients as a penalty, which can shrink some to exactly zero for feature selection; L2 (Ridge) adds the squared value, shrinking coefficients but rarely to zero.
26What is a confusion matrix?
A table showing true positives, true negatives, false positives, and false negatives, used to evaluate a classification model's performance in detail.
27Difference between precision and recall?
Precision measures how many predicted positives were actually correct; recall measures how many actual positives the model successfully identified.
28What is the F1 score, and when would you prioritize it?
The harmonic mean of precision and recall; it's useful when you need a balance between the two, especially on imbalanced datasets.
29Explain the difference between bagging and boosting.
Bagging trains multiple models in parallel on random subsets of data and averages their results to reduce variance; boosting trains models sequentially, each correcting the previous one's errors, to reduce bias.
30What is a Random Forest?
An ensemble of decision trees trained on random subsets of data and features, whose predictions are averaged or voted on to improve accuracy and reduce overfitting.
31How does a decision tree decide where to split?
It selects the split that produces the greatest reduction in impurity, commonly measured using Gini impurity or entropy and information gain.
32What is gradient boosting, and how is it different from Random Forest?
Gradient boosting builds trees sequentially, with each tree correcting the residual errors of the previous ones, unlike Random Forest's parallel, independent trees.
33Explain the curse of dimensionality.
As the number of features grows, data becomes increasingly sparse in the feature space, making distance-based models less effective and increasing the risk of overfitting.
34What is feature scaling, and why does it matter?
It normalizes the range of independent variables so no single feature dominates a distance-based or gradient-based algorithm purely because of its scale.
35What's the difference between a parametric and non-parametric model?
Parametric models assume a fixed functional form with a set number of parameters, like linear regression; non-parametric models grow in complexity with the data, like KNN or decision trees.
36Explain the ROC curve and AUC.
The ROC curve plots the true positive rate against the false positive rate at different thresholds; AUC summarizes overall model performance across all thresholds, with 1.0 being perfect.
37What is k-fold cross-validation?
The dataset is split into k equal parts; the model trains on k-1 folds and validates on the remaining fold, repeating k times so every fold is used for validation once.
38What's the difference between K-Means and hierarchical clustering?
K-Means requires specifying the number of clusters upfront and partitions data iteratively; hierarchical clustering builds a tree of nested clusters without needing a predefined cluster count.
39Explain what hyperparameter tuning is and one common method.
It's the process of finding the best configuration values for a model that aren't learned from data, commonly done using Grid Search or Random Search combined with cross-validation.
40What is data leakage, and how do you prevent it?
It happens when information from outside the training set, often from the future or the test set, improperly influences model training; it's prevented with strict train/test separation and careful feature engineering.
Deep Learning & NLPQ41–Q60
41What is a neural network, in simple terms?
A layered system of interconnected nodes (neurons) that learns to map inputs to outputs by adjusting connection weights through training.
42What is backpropagation?
The algorithm used to train neural networks by calculating the gradient of the loss function with respect to each weight, then updating weights to reduce error.
43What is the vanishing gradient problem?
In deep networks, gradients can shrink exponentially as they're propagated backward through many layers, making earlier layers learn extremely slowly or not at all.
44Explain the difference between CNN and RNN.
CNNs are designed for spatial data like images, using convolutional filters to detect patterns; RNNs are designed for sequential data like text or time series, maintaining a memory of previous inputs.
45What is an activation function, and why is it needed?
It introduces non-linearity into a neural network, allowing it to learn complex patterns instead of behaving like a simple linear model.
46What's the difference between ReLU and Sigmoid activation functions?
ReLU outputs zero for negative inputs and the input itself for positive ones, avoiding vanishing gradients on the positive side; Sigmoid squashes outputs between 0 and 1 but saturates and causes vanishing gradients at extremes.
47What is dropout, and why is it used?
A regularization technique that randomly disables a fraction of neurons during training, forcing the network to avoid relying too heavily on any single neuron, which reduces overfitting.
48Explain what an LSTM is and why it's used over a standard RNN.
Long Short-Term Memory networks use gates to control what information is kept or forgotten, solving the vanishing gradient problem that limits standard RNNs on long sequences.
49What is a transformer architecture?
A model architecture built entirely around self-attention mechanisms rather than recurrence, allowing it to process sequences in parallel and capture long-range dependencies more effectively.
50What is attention in the context of deep learning?
A mechanism that lets a model weigh the importance of different parts of the input when producing each part of the output, rather than treating all inputs equally.
51What is tokenization in NLP?
The process of breaking text into smaller units, like words or subwords, that a model can process numerically.
52Explain the difference between stemming and lemmatization.
Stemming crudely chops word endings to get a root form; lemmatization uses vocabulary and grammar rules to return the actual dictionary base form of a word.
53What are word embeddings?
Dense vector representations of words that capture semantic meaning, so words with similar meanings end up close together in the vector space.
54What's the difference between Word2Vec and TF-IDF?
Word2Vec learns dense semantic embeddings through a neural network trained on context; TF-IDF is a sparse statistical weighting based on word frequency and rarity across documents.
55What is transfer learning?
Reusing a model pretrained on a large dataset and fine-tuning it on a smaller, task-specific dataset, saving training time and often improving performance.
56Explain batch normalization.
A technique that normalizes the inputs to each layer during training, stabilizing and speeding up learning while reducing sensitivity to weight initialization.
57What's the difference between epoch, batch, and iteration in training?
An epoch is one full pass through the training data; a batch is a subset of data processed at once; an iteration is one update step, equal to processing one batch.
58What is an autoencoder used for?
A neural network trained to reconstruct its input, commonly used for dimensionality reduction, anomaly detection, or denoising data.
59Explain what a loss function does.
It quantifies how far the model's predictions are from the actual values, giving the optimizer a signal to adjust weights and improve accuracy.
60What's the difference between a generative and a discriminative model?
A generative model learns the underlying data distribution and can create new samples; a discriminative model focuses only on distinguishing between classes given the input.
Statistics & ProbabilityQ61–Q80
61What is conditional probability, and why does it matter in ML?
The probability of an event occurring given that another event has already happened; it underlies models like Naive Bayes and is central to reasoning about dependent features.
62Explain Bayes' Theorem in simple terms.
It's a way to update the probability of a hypothesis as new evidence becomes available, combining prior belief with the likelihood of the observed data.
63What's the difference between a discrete and a continuous probability distribution?
A discrete distribution deals with countable outcomes, like a dice roll; a continuous distribution deals with outcomes over a continuous range, like height, described using probability density functions.
64What is a normal distribution, and why is it so common in statistics?
A symmetric, bell-shaped distribution where most values cluster around the mean; it's common because many natural and averaged phenomena tend toward it, per the Central Limit Theorem.
65Explain what variance measures.
The average squared deviation of data points from the mean, capturing how spread out a dataset is.
66What is covariance, and how is it different from correlation?
Covariance measures how two variables change together in raw units; correlation normalizes that relationship to a standardized range between -1 and 1, making it easier to interpret.
67What is the law of large numbers?
As a sample size grows, the sample average converges toward the true population average, which is why larger datasets tend to give more reliable estimates.
68What is a Z-score, and what does it represent?
It measures how many standard deviations a data point is from the mean, used to compare values across different distributions or detect outliers.
69Explain the difference between a one-tailed and a two-tailed hypothesis test.
A one-tailed test checks for an effect in a single specified direction; a two-tailed test checks for an effect in either direction.
70What is sampling bias, and how can it affect a model?
It occurs when the sample used doesn't represent the true population, leading a trained model to learn skewed patterns that don't generalize well to real-world data.
71What's the difference between a prior and a posterior probability in Bayesian statistics?
A prior is your belief about a parameter before seeing data; a posterior is the updated belief after incorporating observed evidence via Bayes' Theorem.
72What is entropy in the context of information theory?
A measure of uncertainty or randomness in a variable's distribution; it's used in decision trees to decide the most informative feature to split on.
73Explain what a chi-square test is used for.
It tests whether there's a significant association between two categorical variables by comparing observed frequencies to expected frequencies.
74What's the difference between independent and dependent events?
Independent events don't influence each other's probability; dependent events do, meaning the outcome of one changes the likelihood of the other.
75What does "statistically significant" actually mean?
It means the observed result is unlikely to have occurred by random chance alone, based on a predefined significance threshold, typically p < 0.05.
76What's the difference between population variance and sample variance formulas?
Sample variance divides by n-1 instead of n, known as Bessel's correction, to correct for the bias introduced by estimating the mean from the same sample.
77Explain what a probability density function (PDF) represents.
It describes the relative likelihood of a continuous random variable taking on a specific value, where the area under the curve over an interval gives the probability of falling within that range.
78What's the difference between MLE and MAP estimation?
Maximum Likelihood Estimation finds parameters that maximize the likelihood of the observed data alone; Maximum A Posteriori incorporates a prior belief about the parameters along with that likelihood.
79How would you test if two groups have significantly different means?
Using a t-test if comparing two groups with roughly normal distributions, or a non-parametric alternative like the Mann-Whitney U test if normality assumptions don't hold.
80What's the difference between homoscedasticity and heteroscedasticity in regression?
Homoscedasticity means the residuals have constant variance across all levels of the independent variable; heteroscedasticity means that variance changes, which can violate regression assumptions.
Model Evaluation & DeploymentQ81–Q100
81What is the train-test split, and why is it necessary?
Splitting data into separate training and testing sets lets you evaluate how well a model generalizes to unseen data, instead of just checking how well it memorized the training set.
82What's the difference between a validation set and a test set?
The validation set is used during training to tune hyperparameters and make model decisions; the test set is held back entirely and used only for a final, unbiased performance check.
83What's the difference between MAE and RMSE for regression evaluation?
MAE treats all errors equally; RMSE penalizes larger errors more heavily because of the squaring step, making it more sensitive to outliers.
84What is R-squared, and what does it tell you?
It represents the proportion of variance in the dependent variable that's explained by the model's independent variables, with values closer to 1 indicating a better fit.
85Explain what model drift is.
The gradual degradation of a deployed model's performance over time as real-world data patterns shift away from what the model was originally trained on.
86What is A/B testing used for in model deployment?
Comparing a new model's real-world performance against the current production model by routing a portion of live traffic to each and measuring the outcome.
87What's the difference between online and batch prediction?
Online prediction serves results in real time as requests come in; batch prediction processes a large set of inputs together on a schedule and stores the results.
88What is model versioning, and why does it matter?
Tracking different iterations of a trained model so you can reproduce results, roll back to a previous version, or compare performance across versions.
89Explain what a REST API is used for in model deployment.
It exposes a trained model as a web service so other applications can send input data and receive predictions over HTTP.
90What is the precision-recall tradeoff, and how do you manage it?
Increasing the classification threshold typically raises precision but lowers recall, and vice versa; the right balance depends on whether false positives or false negatives are more costly for the business.
91What is feature importance, and how would you calculate it?
A measure of how much each input feature contributes to a model's predictions, often calculated using built-in tree-based importances, permutation importance, or SHAP values.
92What is SHAP, and why is it used?
A method based on game theory that explains individual predictions by fairly attributing the contribution of each feature to the final output.
93What's the difference between model interpretability and explainability?
Interpretability refers to how naturally understandable a model's internal logic is, like a simple decision tree; explainability refers to techniques used to explain a complex "black box" model's decisions after the fact.
94What is the cold start problem in recommendation systems?
The difficulty of making accurate predictions for new users or items that have little to no historical interaction data.
95How would you monitor a model in production?
Track prediction accuracy against ground truth when available, monitor input data distribution for drift, and set alerts for unusual latency or error rates.
96What's the difference between a pipeline and a single model in ML workflows?
A pipeline bundles preprocessing steps, like scaling or encoding, together with the model into a single reusable object, ensuring the same transformations are applied consistently at training and inference time.
97What is early stopping, and why is it used?
A technique that halts training once validation performance stops improving, preventing the model from overfitting to the training data.
98What is a champion-challenger model setup?
The champion is the current production model; a challenger model is tested alongside it on live or held-out data to see if it should replace the champion.
99How do you handle imbalanced classes in a classification problem?
Techniques include resampling, either oversampling the minority class or undersampling the majority, class weighting, or using metrics like F1 or AUC instead of raw accuracy.
100What would you check first if a deployed model's accuracy suddenly drops?
Check for changes in the input data distribution, upstream data pipeline issues, or schema changes before assuming the model itself has degraded.

Ready to Practice These Live?

Our Data Science program includes real mock interview rounds built around exactly these 100 questions, with mentor feedback after each one.