Feature Scaling

Feature scaling standardizes the range or distribution of features so that no single feature dominates distance calculations.

Ask the Data science assistant 1 min read · Updated September 9, 2026

Definition

Standardization (z-score): $x' = (x - \mu) / \sigma$ , giving features mean 0 and std 1.

Min-max scaling: $x' = (x - x_{\min}) / (x_{\max} - x_{\min})$ , mapping to [0, 1].

Distance-based models (k-NN, SVM, k-means, neural nets) are sensitive to feature scale — a feature with large values dominates Euclidean distance.

Linear models, tree-based models, and probability-based models are generally invariant to scaling — but it doesn't hurt them.

Height (cm) and weight (kg): height ~[150, 200], weight ~[40, 120] — weight has higher variance but not necessarily more predictive power.

Image pixel values [0, 255] are often standardized to roughly [-1, 1] for neural network training.

RobustScaler uses median and IQR instead of mean and std — appropriate when data has outliers.

For highly skewed data, log transform or Box-Cox transform can make distributions more symmetric before scaling.

Always fit scaling on the training set only, then apply the same transformation to test/production data.

In a cross-validation pipeline, scaling must be done inside each fold to avoid leakage.

Assembled from the ReLU.chat curated knowledge base. These explanations are concise on purpose; check the sources for anything important.