Min-Max scaling vs Z-score standardization: differences and algorithm preferences

WHAT THIS TESTS: This question probes whether you can move beyond cookbook preprocessing and explain the statistical mechanics behind two fundamental scaling strategies. Interviewers want to see that you understand how the transformed distribution properties interact with an algorithm's objective function, distance metric, or update rule.

A GOOD ANSWER COVERS: First, define Min-Max scaling as compressing features into a fixed bounded range typically zero to one using the formula X minus X min divided by X max minus X min, and note that this suppresses the effect of outliers because it shrinks standard deviations. Second, define Z-score standardization as centering features at mean zero with standard deviation one using z equals x minus mu divided by sigma, which preserves outlier magnitude and handles varying units. Third, pick an algorithm where one is strongly preferred, such as standardization for PCA because maximizing variance along orthogonal components would otherwise be dominated by variables with larger original scales, or standardization for gradient descent in logistic regression and neural networks because feature values directly modulate weight updates and unequal scales cause some weights to update faster than others.

COMMON WRONG ANSWERS: A major red flag is stating that tree-based methods like decision trees or random forests require feature scaling; these algorithms are scale-invariant because splits depend only on rank ordering, not absolute magnitude. Another weak response is claiming Min-Max and Z-score are interchangeable in all contexts without acknowledging that bounded range versus centering at zero matters for optimization stability and outlier sensitivity.

LIKELY FOLLOW-UPS: The interviewer may ask how you would handle a feature with extreme outliers, which usually pushes you toward standardization or robust scaling instead of Min-Max. They might also ask whether you would scale before or after train-test splitting, which tests your awareness of data leakage, or they could ask about the interaction between scaling and regularization in linear models.

ONE CONCRETE EXAMPLE: Suppose you are building a logistic regression model to predict customer churn using age in years ranging from eighteen to ninety and annual income in dollars ranging from twenty thousand to two hundred thousand. If you use Min-Max scaling, both features occupy zero to one, but the income feature's larger variance and potential outliers compress the age signal. If you use Z-score standardization instead, both features have mean zero and standard deviation one, so gradient descent converges faster because the weight updates for age and income are on comparable scales and no single feature dominates the loss surface.

Source: sebastianraschka.com