Data Normalization

Data Normalization is ambiguous in Machine Learning and Data Science: Feature Scaling—rescaling individual numeric features; Unit Norm Normalization—scaling samples to unit norm; or Database Normalization—structuring relational database tables. Feature scaling includes Min-Max Scaling, which maps feature values to a chosen range such as [0, 1] or [−1, 1], and Standardization, also called Z-Score Normalization, which centers features at zero and scales nonconstant features to unit variance. In scikit-learn, “normalization” specifically means scaling each sample to unit norm, rather than standardizing features. Feature scaling is often beneficial for distance-based methods such as K-Nearest Neighbors, margin- or kernel-based methods such as Support Vector Machines, and optimization using Gradient Descent: it can reduce the disproportionate influence of larger-scale features and improve convergence, but benefits depend on the algorithm, data representation, and architecture. It is not universally required and should not be conflated with internal techniques such as Batch Normalization in Deep Learning models, including Convolutional Neural Networks. Database normalization instead uses Normal Forms to reduce Data Redundancy and improve Data Integrity.