What is normalization?
Normalization is the step of rescaling numeric inputs so they share a comparable range before a model sees them. Min-max scaling maps each feature onto a range from zero to one. Standardization subtracts the mean and divides by the standard deviation, so every feature has a mean of zero and a spread of one. Both are called normalization in everyday use.
The need arises because many algorithms measure distance or apply the same penalty to every input. A distributor’s customer record might hold annual order value in the hundreds of thousands and days since last order in the tens. Without normalization, k-nearest neighbors, support vector machines, neural networks, and any regularized linear model will treat order value as far more important simply because its numbers are bigger. Tree-based models are the main exception and are largely indifferent to scale.
Normalization in practice
The scaling parameters must be computed on the training data only and then applied unchanged to validation, test, and production data. Fitting them on the whole dataset is a form of information leakage. The same parameters must ship with the model, because a production record scaled with different numbers produces wrong predictions without any error message.
In data management the word has a second meaning: organizing database tables to remove duplicated data. The two uses are unrelated, and a data engineer and a data scientist can talk past each other for some time before noticing.