How it works
In ML, features with very different scales can distort learning. For example, "number of purchases" (0-50) and "amount spent" (0-50000). The model may give excessive weight to amount spent just because it is larger.
Min-max normalization puts values between 0 and 1. Standardization (z-score) puts values with mean 0 and standard deviation 1. Both solve the scale problem.
In text AI, normalization can mean standardizing text: converting to lowercase, removing accents, normalizing spaces. This ensures that "Café" and "cafe" are treated the same way.
Practical example
When training a model that predicts customer churn, the feature "number of purchases" (0-50) and "amount spent" (0-50000) need to be normalized. Without normalization, the model may consider "amount spent" more important just because the numbers are larger.
In text processing, normalizing "São Paulo", "são paulo" and "SAO PAULO" to "sao paulo" ensures that the model treats all variations as the same entity.
What it's for
Improve training convergence and model quality. Without normalization, training can be slow or unstable. With normalization, the model learns more efficiently and fairly.
For developers, this means including normalization in the data pipeline. For data scientists, it means choosing the right technique for each type of data.
Difference between normalization and standardization
Normalization (min-max) puts values in a fixed interval, usually 0 to 1. Standardization (z-score) puts values with mean 0 and standard deviation 1, with no fixed limit. Use normalization when you need bounded values. Use standardization when the data has outliers or a non-uniform distribution.