✅ Data Science Interview Questions with Answers Part-3

21. How do you handle missing values?
Missing values are handled based on the reason and the impact on the problem. You first check whether data is missing at random or systematic. Common approaches include removing rows or columns if the missing percentage is small, imputing with mean, median, or mode for numerical data, using a separate category for missing values in categorical data, or applying model-based imputation when data loss affects predictions.

22. How do you treat outliers?
Outliers are treated after understanding their cause. If they result from data entry errors, they are corrected or removed. If they represent real but rare events, they are kept. Treatment methods include capping values, applying transformations like log scaling, or using robust models that handle outliers naturally. Blind removal is avoided.

23. What is data normalization and standardization?
Normalization rescales data to a fixed range, usually between zero and one. Standardization rescales data to have a mean of zero and a standard deviation of one. Both techniques ensure features contribute equally to model learning, especially for distance-based and gradient-based algorithms.

24. When do you use Min-Max scaling vs Z-score?
Min-Max scaling is used when data has a fixed range and no extreme outliers, such as image pixel values. Z-score scaling is used when data follows a normal distribution or contains outliers. Many machine learning models perform better with standardized data.

25. How do you handle imbalanced datasets?
Imbalanced datasets are handled by resampling techniques like oversampling the minority class or undersampling the majority class. You can also use algorithms that support class weighting or focus on metrics like recall, precision, and AUC instead of accuracy. The choice depends on business cost of false positives and false negatives.

26. What is one-hot encoding?
One-hot encoding converts categorical variables into binary columns. Each category becomes a separate column with values zero or one. This avoids ordinal assumptions and works well with most machine learning algorithms, especially linear and tree-based models.

27. What is label encoding?
Label encoding assigns a unique numeric value to each category. It is suitable when categories have an inherent order or when using tree-based models that handle ordinal values well. It is avoided for nominal data in linear models due to unintended ranking.

28. How do you detect data leakage?
Data leakage is detected by checking whether future or target-related information is present in training features. You validate time-based splits, review feature creation logic, and ensure preprocessing steps are applied separately on training and test data. Sudden high model accuracy is often a red flag.

29. What is duplicate data and how do you handle it?
Duplicate data refers to repeated records representing the same entity or event. Duplicates are identified using unique identifiers or key feature combinations. They are removed or merged based on business logic to prevent bias, inflated metrics, and incorrect model learning.

30. How do you validate data quality?
Data quality is validated by checking completeness, consistency, accuracy, and validity. This includes range checks, schema validation, distribution analysis, and reconciliation with source systems. Automated checks and dashboards are often used to monitor quality continuously.

Double Tap ♥️ For Part-4