Definition
Normalization often maps values to a fixed range (e.g., 0–1). Standardization centers features to mean 0 and variance 1. Both help optimizers and distance-based methods.
Tree ensembles are less sensitive to scale; neural nets and k-NN usually care a lot.
In simple terms
Comparing test scores from different grading scales — convert them to percentiles before averaging.
Where you see it
- Scaling sensor readings before a neural net.
- Standardizing embeddings is less common; scaling tabular inputs is routine.
How it works
1.Fit on train only
Learn min/max or mean/std from training data.
2.Transform
Apply the same scaler to validation and production.
3.Avoid leakage
Never fit scalers on the full dataset including test.
Why it matters
- Proper scaling stabilizes training and makes features comparable.
Often confused
Normalize after splitting randomly each time in production.
Persist training scaler parameters and reuse them.