Company-Wise Data Scientist Interview Questions: The Ultimate Q&A [ Part - 1 ]👨‍⚖️

Crack Your Data Science Interview: Top Questions & Answers

Are you preparing for a Data Scientist role? You've come to the right place! Cracking a data science interview requires a solid understanding of fundamental concepts, from statistical theorems to machine learning algorithms and practical implementation. This guide compiles frequently asked interview questions from top companies like L&T, Tiger Analytics, Infosys, Siemens, Wipro, and Deloitte. We've structured this post as an interactive Q&A to help you learn, revise, and ace your next interview.

Quick Navigation: Interview Questions By Company

• 🏢 L&T Financial Services
• 🐯 Tiger Analytics
• 💻 Infosys
• ⚡ Siemens
• 🌐 Wipro
• 🔵 Deloitte

🏢 Company: L&T Financial Services

Role: Data Scientist

1. 1. Explain your Projects

   Question: Can you walk me through one of your most significant data science projects?

   This is the most common opening question. Structure your answer using the STAR method (Situation, Task, Action, Result).

   • Situation: Describe the business problem. (e.g., "The company was facing a high customer churn rate of 15%...")
   • Task: What was your goal? (e.g., "My task was to build a predictive model to identify customers at high risk of churning.")
   • Action: What did you do? (e.g., "I engineered features like customer tenure and usage patterns. I trained several models, including Logistic Regression and a Random Forest, and found the Random Forest performed better.")
   • Result: What was the outcome? Quantify your impact. (e.g., "The final model achieved an accuracy of 88% and helped the marketing team target at-risk customers, leading to a 3% reduction in churn in the next quarter.")

   Pro-Tip: Always have 2-3 projects ready to discuss. Tailor the project you lead with to the company's industry. For a financial services company, a project on fraud detection or credit risk scoring is ideal.
2. 2. Assumptions in Multiple linear regression

   Question: What are the key assumptions a dataset must meet to use Multiple Linear Regression effectively?

   Multiple Linear Regression relies on several key assumptions:

   1. Linearity: The relationship between independent and dependent variables is linear.
   2. Independence: The residuals (errors) are independent. No autocorrelation.
   3. Homoscedasticity: The variance of residuals is constant across all levels of independent variables.
   4. Normality of Residuals: The residuals are approximately normally distributed.
   5. No Multicollinearity: The independent variables are not highly correlated with each other.
3. 3. Decision tree algorithm

   Question: Can you give me a high-level explanation of how a Decision Tree algorithm works?

   A Decision Tree is a supervised learning algorithm that works by splitting data into subsets based on feature values. It's a tree-like model where each internal node represents a test on a feature, each branch is an outcome of the test, and each leaf node is a class label (classification) or a continuous value (regression).

4. 4. Gini index

   Question: What is the Gini Index, and how is it used in decision trees?

   The Gini Index (or Gini Impurity) is a metric used by decision trees to measure the impurity of a node. It calculates the probability of misclassifying a randomly chosen element from the node. The algorithm chooses the split that results in the lowest Gini Index for the child nodes.

5. 5. Entropy

   Question: What is Entropy in the context of machine learning, and how does it relate to Information Gain?

   Entropy is another measure of impurity or disorder in a node. A split is chosen based on which one provides the highest Information Gain, which is simply the reduction in entropy after the split.

6. 6. Formulas of Gini and entropy

   Question: Can you write down the formulas for Gini Impurity and Entropy?

   Gini Formula: G = 1 - Σ (pᵢ)²
   Entropy Formula: E = - Σ (pᵢ * log₂(pᵢ))
   Where pᵢ is the probability of an element belonging to class i in the node.
7. 7. Random forest algorithm

   Question: How does a Random Forest improve upon a single Decision Tree?

   A Random Forest is a bagging ensemble method. It builds multiple decision trees on random subsets of data (bootstrapping) and random subsets of features. By averaging their predictions, it reduces the high variance of a single decision tree, leading to less overfitting and a more robust model.

8. 8. XGBoost Algorithm

   Question: What makes XGBoost so popular and powerful?

   XGBoost (Extreme Gradient Boosting) is a boosting algorithm. It builds trees sequentially, with each new tree correcting the errors of the previous ones. Its power comes from its speed (parallel processing), built-in regularization to prevent overfitting, and its ability to handle missing values automatically.

9. 9. Central Limit theorem

   Question: Can you explain the Central Limit Theorem and why it's so important in statistics?

   The Central Limit Theorem (CLT) states that the sampling distribution of the sample mean will be approximately normally distributed, regardless of the original population's distribution, as long as the sample size is large enough (n > 30). It's crucial because it allows us to perform hypothesis tests and create confidence intervals even when we don't know the population's distribution.

10. 10. R2

    Question: What does the R-squared metric tell you about a regression model?

    R-squared (R²) represents the proportion of the variance in the dependent variable that is predictable from the independent variable(s). It's a measure of how well the model's predictions fit the actual data, with a value of 1 indicating a perfect fit.

11. 11. Adj R2

    Question: Why would you use Adjusted R-squared instead of R-squared?

    Adjusted R-squared is used when comparing models with different numbers of features. R² will always increase if you add more features, even if they are useless. Adjusted R² only increases if the new feature improves the model more than would be expected by chance, making it a better metric for model selection.

12. 12. VIF

    Question: What is VIF and how do you use it?

    Variance Inflation Factor (VIF) is a metric used to detect multicollinearity. It measures how much the variance of a regression coefficient is inflated due to its correlation with other predictors. A VIF greater than 5 or 10 is a common threshold to indicate problematic multicollinearity.

13. 13. Different Methods to measure Accuracy

    Question: Beyond accuracy, what other metrics would you use to evaluate a model's performance?

    Accuracy isn't always the best metric, especially for imbalanced datasets. Other key metrics include:

    • Classification: Precision, Recall, F1-Score, AUC-ROC Curve, Log-Loss, Confusion Matrix.
    • Regression: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), R-squared.
14. 14. Explain Bagging and Boosting

    Question: Can you briefly explain the concepts of Bagging and Boosting?

    Both are ensemble techniques. Bagging trains models in parallel on random data subsets to reduce variance (e.g., Random Forest). Boosting trains models sequentially, where each model corrects the errors of the previous one, to reduce bias (e.g., XGBoost).

15. 15. Difference Between Bagging and Boosting

    Question: What are the key differences between Bagging and Boosting?

    Bagging (e.g., Random Forest)

    • Parallel: Trees are built independently.
    • Bootstrap Sampling: Trains on random data subsets.
    • Goal: Reduce variance.
    • Voting: Final prediction by average/vote.

    Boosting (e.g., XGBoost)

    • Sequential: Trees are built one after another.
    • Weighted Data: Focuses on previous errors.
    • Goal: Reduce bias.
    • Weighted Sum: Final prediction is a weighted sum.

    Key differences between Bagging and Boosting.

16. 16. Various Ensemble techniques

    Question: Besides Bagging and Boosting, what other ensemble techniques are you aware of?

    The main types are:

    • Bagging: Reduces variance (e.g., Random Forest).
    • Boosting: Reduces bias (e.g., XGBoost).
    • Stacking (or Blending): Trains multiple different models and uses a meta-model to combine their predictions.
17. 17. P-value and its significance

    Question: In statistical testing, what does a p-value represent?

    The p-value is the probability of observing data as extreme as, or more extreme than, what was actually observed, assuming the null hypothesis is true. A small p-value (typically ≤ 0.05) provides evidence to reject the null hypothesis.

18. 18. F1 Score

    Question: When would you use the F1 Score, and what does it measure?

    The F1 Score is the harmonic mean of Precision and Recall. It is a great metric for imbalanced datasets because it seeks a balance between Precision (not making false positive errors) and Recall (not making false negative errors).

    F1 Score = 2 * (Precision * Recall) / (Precision + Recall)
19. 19. Type 1 and Type II error

    Question: Can you explain the difference between a Type I and a Type II error?

    In hypothesis testing:

    • Type I Error (False Positive): Rejecting a true null hypothesis.
    • Type II Error (False Negative): Failing to reject a false null hypothesis.
20. 20. Logical questions for Type 1 and Type II error

    Question: In the context of testing a new drug, what would a Type I and Type II error represent?

    Scenario: Null Hypothesis (H₀) is "The drug has no effect."

    • Type I Error: Concluding the drug is effective when it is not (False Positive). Patients take a useless drug.
    • Type II Error: Concluding the drug is not effective when it actually is (False Negative). A potentially life-saving drug is discarded.
21. 21. Logical questions for Null and alternate Hypothesis

    Question: If we are testing a new website design, how would you formulate the Null and Alternate Hypotheses?

    Scenario: Testing if a new design increases user engagement time.

    • Null Hypothesis (H₀): The new design has no effect or decreases the average engagement time. (μ_new ≤ μ_old)
    • Alternate Hypothesis (H₁): The new design increases the average engagement time. (μ_new > μ_old)

🐯 Company: Tiger Analytics

Role: Data Scientist

1. 1. You are given a data set with missing values that spread along 1 standard deviation from the median. What percentage of data would remain unaffected?

   Question: You are given a data set with missing values that spread along 1 standard deviation from the median. What percentage of data would remain unaffected?

   This tests your understanding of the Normal Distribution and the Empirical Rule (68-95-99.7). Assuming the data is roughly normal, the mean and median are close. One standard deviation (1σ) from the median covers about 68% of the data.

   Therefore, the percentage of data that would remain unaffected is:

   100% - 68% = 32%

   Pro-Tip: State your assumption: "Assuming a near-normal distribution where the median approximates the mean, about 32% of the data would lie outside one standard deviation."
2. 2. Explain the difference between an array and a linked list.

   Question: What are the main differences between an array and a linked list in terms of memory and performance?

   Both are linear data structures, but:

   • Array: Stores elements in a contiguous block of memory. Fast O(1) random access, but slow O(n) insertion/deletion.
   • Linked List: Stores elements non-contiguously using pointers. Slow O(n) access, but fast O(1) insertion/deletion.
3. 3. How do you ensure you are not overfitting a model?

   Question: What techniques do you use to prevent a model from overfitting?

   Overfitting is when a model learns training data noise. To prevent it:

   • Cross-Validation: Use k-fold cross-validation for robust performance estimation.
   • Simplify the Model: Use a less complex model or fewer features.
   • Regularization: Use L1 (Lasso) or L2 (Ridge) to penalize complexity.
   • Pruning / Dropout: Techniques for trees and neural networks, respectively.
   • Get More Data: More data helps the model learn the true signal.
4. 4. How do you fix high variance in a model?

   Question: If your model is suffering from high variance, what steps would you take to address it?

   High variance is a synonym for overfitting. The solutions are:

   • Increase Training Data: The most effective solution.
   • Use Bagging: Ensemble methods like Random Forest are designed to reduce variance.
   • Apply Regularization: L1 or L2 regularization constrains model complexity.
   • Reduce Model Complexity: Use fewer features or a simpler algorithm.
5. 5. What are hyperparameters? How do they differ from model parameters?

   Question: Can you explain the difference between a model's hyperparameters and its parameters?

   The difference is about who sets the value and when.

   • Hyperparameters: Set before training begins by the data scientist (e.g., learning rate, k in KNN). They control the learning process.
   • Model Parameters: Learned during training from the data (e.g., weights in a neural network, coefficients in a linear regression).
6. 6. What is the default method for splitting in decision trees? What other methods are available?

   Question: What criteria do decision trees use to decide on a split, and are there different options?

   The goal is to maximize purity in child nodes.

   • Default Methods: Gini Index (used by CART) and Information Gain (using Entropy, used by ID3/C4.5). Scikit-learn defaults to Gini for classification.
   • Other Methods: For regression, the criterion is Variance Reduction (or Mean Squared Error).
7. 7. You are told that your regression model is suffering from multicollinearity. How do verify this is true and build a better model?

   Question: Imagine your regression model might have multicollinearity. How would you confirm it, and what steps would you take to fix it?

   Verification:

   1. Correlation Matrix: A quick visual check for highly correlated pairs.
   2. Variance Inflation Factor (VIF): The standard method. A VIF > 5-10 indicates a problem.

   Building a Better Model:

   1. Remove a Variable: Remove one of the correlated features.
   2. Use Regularization: Ridge Regression (L2) is very effective at handling multicollinearity.
   3. Use PCA: Transform variables into uncorrelated components.
8. 8. You build a random forest model with 10,000 trees. Training error is at 0.00, but the validation error is 34.23. Explain what went wrong.

   Question: I've built a random forest with 10,000 trees. The training error is zero, but validation error is very high. What's happening?

   This is a classic case of severe overfitting. The model has memorized the training data perfectly but has completely failed to generalize to unseen data.

   Cause: Extreme model complexity. The trees are likely too deep and numerous. Solution: Tune the hyperparameters. Drastically reduce `n_estimators` (number of trees) and set limits like `max_depth` and `min_samples_leaf` to simplify the trees.

9. 9. What is the recall, specificity, and precision of the confusion matrix?

   Question: Can you define Precision, Recall, and Specificity using the components of a confusion matrix?

   Predicted Positive

   Predicted Negative

   Actual
   Positive

   TP

   FN

   Actual
   Negative

   FP

   TN

   Structure of a Confusion Matrix.

   • Precision: Of all positive predictions, how many were correct?
     Precision = TP / (TP + FP)
   • Recall (Sensitivity): Of all actual positives, how many did the model find?
     Recall = TP / (TP + FN)
   • Specificity: Of all actual negatives, how many did the model find?
     Specificity = TN / (TN + FP)

💻 Company: Infosys

Role: Data scientist

1. 1) curse of dimensionality? How would you handle it?

   Question: What is the "curse of dimensionality," and what are your main strategies for handling it?

   The Curse of Dimensionality refers to problems arising from high-dimensional data, where data becomes sparse, distances become less meaningful, and overfitting risk increases.

   How to Handle It:

   • Feature Selection: Use methods to select only the most relevant features (e.g., using correlation or L1 Regularization).
   • Dimensionality Reduction: Use techniques like PCA to transform features into a smaller set of components.
2. 2) How to find the multicollinearity in the data set

   Question: What methods would you use to detect multicollinearity in a dataset?

   The two main methods are:

   1. Correlation Matrix with a Heatmap: A visual check for high correlations (>|0.8|) between independent variables.
   2. Variance Inflation Factor (VIF): A more definitive test. VIF > 5 or 10 indicates significant multicollinearity.
3. 3)Explain the different ways to treat multicollinearity!

   Question: Once you've found multicollinearity, how would you go about treating it?

   Several strategies work well:

   • Remove one of the correlated features.
   • Combine the features into a single new feature.
   • Use a model with built-in regularization, like Ridge Regression.
   • Use Principal Component Analysis (PCA) to create uncorrelated components.
4. 4) How do you decide which feature to keep and which feature to eliminate after performing the multicollinearity test?

   Question: When two features are highly correlated, how do you decide which one to drop?

   The decision combines statistics and business sense:

   • Check VIF Scores: Drop the one with the higher VIF.
   • Correlation with Target: Keep the one with a stronger correlation to the target variable.
   • Domain Knowledge: Keep the feature that is more important or interpretable for the business.
   • Data Completeness: Consider dropping the feature with more missing values.
5. 5)Explain logistic regression

   Question: Can you give me a concise explanation of Logistic Regression?

   Logistic Regression is a supervised algorithm for binary classification. It calculates a weighted sum of inputs and passes it through a Sigmoid function, which maps the output to a probability between 0 and 1. A threshold (like 0.5) is used to assign the final class.

6. 6)we have a sigmoid function which gives us the probabilty between 0-1 then what is the need for log loss in logistic regression?

   Question: If the sigmoid function already gives a probability, why do we need the Log Loss function in logistic regression?

   While sigmoid provides the output, we need a cost function to measure the model's error during training. Log Loss is used because it heavily penalizes predictions that are both confident and wrong, which makes it an excellent guide for the model to learn the correct parameters via gradient descent.

7. 7) P-value and its significance in statistical testing?

   Question: What is a p-value and what is its significance in hypothesis testing?

   The p-value is the probability of observing data as extreme as, or more extreme than, the current observation, assuming the null hypothesis is true. Its significance lies in decision-making: if the p-value is below a chosen threshold (alpha, e.g., 0.05), we have enough evidence to reject the null hypothesis.

8. 8) How do you split the time series data and evaluation metrics for time series data

   Question: For a time series forecasting problem, how would you split your data for validation, and what metrics would you use?

   Splitting Time Series Data

   You must use a chronological split to avoid data leakage. Random splits are incorrect. Options include:

   • Train-Test Split: Train on older data, test on recent data (e.g., train on 2020-2022, test on 2023).
   • Walk-Forward Validation: An iterative approach where you train, test, then add the test data to the training set for the next iteration.

   Evaluation Metrics

   • MAE (Mean Absolute Error): Easy to interpret.
   • RMSE (Root Mean Squared Error): Penalizes large errors more.
   • MAPE (Mean Absolute Percentage Error): Expresses error as a percentage.
9. 9) How did you deploy your model in production? How often do you retrain it?

   Question: Can you describe the process of deploying a model into production and your strategy for retraining it?

   Deployment Process:

   1. Model Serialization: Save the model with `pickle` or `joblib`.
   2. API Wrapper: Create an API endpoint using Flask or FastAPI.
   3. Containerization: Package the app with Docker.
   4. Cloud Deployment: Deploy the container to a service like AWS SageMaker.

   Retraining Strategy:

   • Scheduled: Retrain on a fixed schedule (daily, weekly).
   • Trigger-based: Monitor model performance (data drift) and retrain automatically when performance degrades below a threshold. This is the more advanced MLOps approach.

⚡ Company: Siemens

Role: Data Scientist

1. 1. What graphs do you use for basic EDA?

   Question: What are your go-to visualizations for initial Exploratory Data Analysis?

   The choice depends on the variable type:

   Univariate Analysis (One Variable)

   • Categorical: Bar Chart or Pie Chart.
   • Numerical: Histogram and Box Plot.

   Bivariate Analysis (Two Variables)

   • Numerical vs. Numerical: Scatter Plot.
   • Numerical vs. Categorical: Box Plot or Violin Plot.
   • Categorical vs. Categorical: Heatmap.
2. 2. How to join tables in python?

   Question: In Python, what is the standard way to perform SQL-like joins on data tables?

   The standard way is to use the pandas library, specifically the `pandas.merge()` function, which can perform inner, outer, left, and right joins.

   import pandas as pd
   # Assuming df1 and df2 are pandas DataFrames
   merged_df = pd.merge(df1, df2, on='common_column', how='inner')
   

3. 3. What is the benefit of shuffling a training dataset when using a batch gradient descent algorithm for optimizing a neural network? 

   Question: Is there a benefit to shuffling the training data when using Batch Gradient Descent?

   This is a trick question. For **true Batch Gradient Descent**, the entire dataset is used in every step, so shuffling provides **no benefit** as the gradient calculation is identical regardless of order.

   However, shuffling is **critical** for Stochastic Gradient Descent (SGD) and Mini-batch Gradient Descent to prevent the model from learning order-based patterns and to ensure robust convergence.

4. 4. Why it is not advisable to use a softmax output activation function in a multi-label classification problem for a one-hot encoded target?

   Question: Why is softmax the wrong activation function for a multi-label classification problem?

   Softmax outputs a probability distribution where all outputs sum to 1. This is perfect for **multi-class** problems where an instance belongs to exactly one class. However, in a **multi-label** problem, an instance can have multiple labels simultaneously (e.g., a movie is 'action' AND 'comedy').

   The correct function for multi-label is Sigmoid on each output neuron, as it calculates the probability for each label independently of the others.

5. 5. How can you iterate over a list and also retrieve element indices at the same time?

   Question: In Python, what's the best way to loop over a list while getting both the index and the value of each element?

   The most Pythonic way is to use the built-in enumerate() function.

   my_list = ['a', 'b', 'c']
   for index, value in enumerate(my_list):
       print(f"Index: {index}, Value: {value}")
   

6. 6. For a given dataset, you decide to use SVM as the main classifier. You select RBF as your kernel. What would be the optimum gamma value that would allow you to capture the features of the dataset really well?

   Question: In an SVM with an RBF kernel, how do you determine the optimal value for the gamma hyperparameter?

   There is no single "optimum" value. The gamma parameter controls the influence of a single training example and must be tuned for each specific dataset. The standard method is to use a search algorithm like GridSearchCV or RandomizedSearchCV with cross-validation to test a range of gamma values and select the one that yields the best model performance.

7. 7. What does the cost parameter in SVM stand for?

   Question: What does the 'C' parameter represent in an SVM?

   The C parameter is the regularization parameter. It controls the trade-off between maximizing the margin and minimizing the classification error on the training data. A **low C** value creates a wider margin but may misclassify more points (higher bias, lower variance). A **high C** value creates a smaller margin and tries to classify all points correctly, which can lead to overfitting (lower bias, higher variance).

8. 8. What is Softmax Function? What is the formula of Softmax Normalization?

   Question: Can you explain the Softmax function and provide its formula?

   The Softmax function is an activation function that converts a vector of numbers (logits) into a probability distribution. It's used in the output layer of multi-class classification networks.

   The formula for the i-th element of the output is:

   Softmax(z)ᵢ = e^zᵢ / Σⱼ e^zⱼ

   This ensures each output is between 0 and 1, and all outputs sum to 1.

🌐 Company: Wipro

Role: Data Scientist

1. 1. Difference between WHERE and HAVING in SQL

   Question: What is the difference between the WHERE and HAVING clauses in SQL?

   The key difference is timing:

   • WHERE Clause: Filters rows before grouping. It operates on individual rows.
   • HAVING Clause: Filters groups after `GROUP BY` and aggregation. It operates on aggregated results and can use functions like `COUNT()` or `SUM()`.
2. 2. Basics of Logistics Regression

   Question: Can you cover the basics of Logistic Regression?

   Logistic Regression is a fundamental algorithm for binary classification. It uses the Sigmoid function to map a linear combination of inputs to a probability (0 to 1). A threshold (usually 0.5) is then used to make the final class prediction.

3. 3. How do you treat outliers?

   Question: What is your strategy for handling outliers in a dataset?

   The strategy depends on the context, but common methods include:

   • Removal: If it's a clear error and the number is small.
   • Transformation: Using log or square root transforms to reduce their effect.
   • Imputation: Treating it as a missing value and imputing it.
   • Capping/Winsorizing: Limiting the value to a certain percentile.
   • Use a Robust Model: Algorithms like Random Forest are less sensitive to outliers.
4. 4. Explain the confusion matrix?

   Question: Can you explain the components of a confusion matrix?

   A Confusion Matrix evaluates a classification model by showing:

   • True Positives (TP): Correctly predicted positive cases.
   • True Negatives (TN): Correctly predicted negative cases.
   • False Positives (FP): Type I Error.
   • False Negatives (FN): Type II Error.
5. 5. Explain PCA (Wanted me to explain the covariance matrix and eigenvectors and values and the mathematical expression and mathematical derivation for co-variance matrix)

   Question: Can you explain the mathematical steps behind Principal Component Analysis (PCA)?

   PCA is a dimensionality reduction technique. The steps are:

   1. Standardize the Data: Scale features to have a mean of 0 and standard deviation of 1.
   2. Compute the Covariance Matrix: This matrix shows how the variables vary with respect to each other.
   3. Calculate Eigenvectors and Eigenvalues: Decompose the covariance matrix. Eigenvectors represent the directions of the new feature space (the principal components), and Eigenvalues represent the amount of variance captured by each eigenvector.
   4. Select Principal Components: Rank eigenvectors by their eigenvalues and keep the top 'k' to reduce dimensionality.
   5. Transform the Data: Project the original data onto the selected eigenvectors.
6. 6. How do you cut a cake into 8 equal parts using only 3 straight cuts?

   Question: How do you cut a cake into 8 equal parts using only 3 straight cuts?

   This is a classic lateral thinking puzzle.

   1. Cuts 1 & 2: Two perpendicular cuts through the top center, creating 4 equal pieces.
   2. Cut 3: A single horizontal cut through the middle of the cake's height, slicing all 4 pieces in half.

   This results in 8 equal pieces.

7. 7. Explain K-means clustering

   Question: How does the K-Means clustering algorithm work?

   K-Means is an unsupervised algorithm that partitions data into 'k' clusters.

   1. Initialization: Randomly select 'k' initial centroids.
   2. Assignment Step: Assign each data point to the nearest centroid.
   3. Update Step: Recalculate each centroid as the mean of the points assigned to it.
   4. Repeat: Repeat the assignment and update steps until the centroids stop moving.
8. 8. How is KNN different from k-means clustering?

   Question: What is the difference between K-Nearest Neighbors (KNN) and K-Means Clustering?

   They sound similar but are very different:

   • Algorithm Type: KNN is a supervised algorithm for classification/regression. K-Means is an unsupervised algorithm for clustering.
   • Goal: KNN predicts a label for a new point based on its neighbors. K-Means groups unlabeled data into clusters.
   • Training: KNN is a "lazy learner" (no training phase). K-Means has a training phase to find centroids.
9. 9. What would be your strategy to handle a situation indicating an imbalanced dataset?

   Question: What is your strategy for handling an imbalanced dataset?

   An imbalanced dataset requires special handling:

   • Use Appropriate Metrics: Use Precision, Recall, F1-Score, and AUC-ROC, not accuracy.
   • Resampling: Use oversampling (e.g., SMOTE) to increase the minority class or undersampling to decrease the majority class.
   • Use Class Weights: Penalize errors on the minority class more heavily during training.
   • Use Different Algorithms: Tree-based models often handle imbalance better.
10. 10. Stock market prediction: You would like to predict whether or not a certain company will declare bankruptcy within the next 7 days... Would you treat this as a classification or a regression problem?

    Question: If you need to predict whether or not a company will go bankrupt, is that a classification or a regression problem?

    This is a **binary classification problem**. The target variable is discrete with two outcomes: 'Bankruptcy' (1) or 'No Bankruptcy' (0). A regression problem would involve predicting a continuous value, like the company's future stock price.

🔵 Company: Deloitte

Role: Data Scientist

1. 1. difference between Correlation and Regression.

   Question: Can you explain the difference between Correlation and Regression?

   While both describe relationships, they serve different purposes.

   • Correlation: Measures the strength and direction of a linear relationship. It's a single value (-1 to 1) and doesn't imply causation.
   • Regression: Aims to predict a dependent variable from independent variables. It gives an equation and defines dependency.
2. 2. Why do we square the residuals instead of using modulus?

   Question: In linear regression, why do we minimize the sum of squared errors instead of the sum of absolute errors?

   This is the basis of Ordinary Least Squares (OLS). Reasons include:

   1. Differentiability: The squared error function is smooth and easy to solve mathematically using calculus. The absolute error function is not.
   2. Penalizes Larger Errors More: Squaring errors gives much more weight to large mistakes, forcing the model to avoid them.

   Pro-Tip: Mention that using the absolute error (modulus) is a valid method called **Least Absolute Deviations (LAD)**, which is more robust to outliers.
3. 3. Which evaluation metric should you prefer to use for a dataset having a lot of outliers in it?

   Question: For a regression problem with many outliers, which evaluation metric is most appropriate?

   Metrics that square errors (like MSE or RMSE) are heavily influenced by outliers. The best choice is Mean Absolute Error (MAE). Since it doesn't square the difference, it's less sensitive to the large errors from outliers and gives a more robust measure of average performance.

4. 4. Heteroscedasticity? How to detect it?

   Question: What is heteroscedasticity, and how would you detect it in a regression model?

   Heteroscedasticity is when the variance of the model's errors is not constant. It violates a key assumption of linear regression.

   How to Detect It:

   1. Residual Plots: Plot predicted values vs. residuals. A random cloud of points is good. A pattern (like a cone shape) indicates heteroscedasticity.
   2. Statistical Tests: Use formal tests like the Breusch-Pagan test.
5. 5. p-value?

   Question: What is a p-value?

   The p-value is the probability of obtaining test results at least as extreme as the results actually observed, under the assumption that the null hypothesis is correct. It's a measure of evidence against the null hypothesis.

6. 6. deal with missing random values from a data set?

   Question: What are some common methods for handling randomly missing values in a dataset?

   Strategies for Missing Completely At Random (MCAR) values include:

   • Mean/Median/Mode Imputation: Simple but can distort variance.
   • Regression Imputation: Build a model to predict the missing values.
   • K-Nearest Neighbors (KNN) Imputation: Use the values of the 'k' most similar data points.
   • Use Models that Handle Missing Values: Algorithms like XGBoost can handle them internally.
7. 7. Z test, F test, and T-test?

   Question: Can you explain the difference between a Z-test, T-test, and F-test?

   • T-test: Compares the means of **two** groups when the sample size is small (< 30) and population variance is unknown.
   • Z-test: Compares the means of two groups when the sample size is large (> 30) and population variance is known.
   • F-test: Compares the variances of **two or more** groups. It's the core of ANOVA.
8. 8. Root Cause Analysis?

   Question: What do you understand by Root Cause Analysis?

   Root Cause Analysis (RCA) is a problem-solving method to find the fundamental cause of an issue, rather than just treating the symptoms. Common techniques include the **5 Whys** and the **Fishbone (Ishikawa) Diagram**.

9. 9. lists, sets, and tuples? difference?

   Question: What are the key differences between lists, tuples, and sets in Python?

   • List: Ordered, mutable (changeable), allows duplicates. `[]`
   • Tuple: Ordered, immutable (unchangeable), allows duplicates. `()`
   • Set: Unordered, mutable, does **not** allow duplicates. Optimized for membership tests. `{}`
10. 10. lambda functions? Write small example python code using it.

    Question: What is a lambda function in Python, and can you provide a simple example?

    A lambda function is a small, anonymous function defined with the `lambda` keyword. It can have multiple arguments but only one expression.

    # A lambda function to multiply two numbers
    multiply = lambda a, b: a * b
    print(multiply(7, 8))  # Output: 56
    

11. 11. Regularization?

    Question: What is regularization in machine learning and why is it used?

    Regularization is a technique to prevent overfitting by adding a penalty term to the model's loss function, which discourages large coefficient values.

    • L1 Regularization (Lasso): Can shrink coefficients to exactly zero, performing feature selection.
    • L2 Regularization (Ridge): Shrinks coefficients close to zero, great for multicollinearity.
12. 12. DBSCAN Clustering?

    Question: Can you explain how DBSCAN clustering works and what its advantages are over K-Means?

    DBSCAN is a density-based clustering algorithm. Its advantages over K-Means are that it does **not** require you to specify the number of clusters beforehand, it can find arbitrarily shaped clusters, and it can identify points as noise/outliers.

Final Words & Good Luck!

Preparing for data science interviews is a marathon, not a sprint. This guide covers many of the core concepts you'll face, but the most important thing is to understand the "why" behind each technique. Keep practicing, stay curious, and be confident in your skills. You've got this!

The data science community thrives on sharing knowledge. If this guide helped you, please consider paying it forward by sharing it with colleagues and friends who are also on their interview journey. A simple share can make a huge difference!