A loan model uses annual income (tens of thousands of dollars) and number of open credit cards (usually under 10). Fed in raw, the income column’s huge numbers can dominate distance calculations and slow down training, simply because of its units. Scaling fixes that. Min-max scaling squeezes each column to 0–1. Z-score standardization subtracts the column’s average and divides by its spread, so a value of +2 means “two standard deviations above average.” Log scaling helps for long-tailed values like income, where a few huge numbers would otherwise squash everyone else together.
Google’s guide notes models usually train faster and predict better when numeric features share a similar range. It matters most for methods that rely on gradients or distances, such as neural networks and linear or logistic regression, and much less for tree models such as a random forest.
The common trap is computing the scaling from the whole dataset. The averages and ranges must be learned from the training set only and then applied unchanged to validation, test and production data; otherwise information about the test data leaks into training (data leakage). “Normalization” also names unrelated ideas elsewhere, such as layers inside a neural network that normalize activations, or database normalization, which is why this entry autolinks only the feature-scaling phrases.
Sources
- Google Machine Learning Crash Course: Numerical data: Normalization — Describes linear scaling, Z-score scaling, log scaling and clipping as normalization techniques suited to different distributions; checked October 24, 2026.
- Google Machine Learning Glossary: normalization — Defines normalization as converting a variable's actual range into a standard range such as -1 to +1, 0 to 1 or Z-scores, and says models usually train faster when numeric features share a similar range; checked October 24, 2026.
- scikit-learn: Preprocessing data — Documents standardization, min-max and robust scaling, and fitting the scaler on training data before applying it to new data; checked October 24, 2026.