A maintenance team combines machine temperature with operating hours for a prediction model. The features use different units, so the team checks whether their scales affect training.
The scikit-learn preprocessing guide describes standardization and min-max scaling among its preprocessing methods. The guide also describes a scaler that uses statistics less affected by outliers.
Standardization centers a feature using its training mean and scales it using its training standard deviation. Min-max scaling uses the training minimum and maximum to map values into a chosen range; values outside the training range can fall outside that range later.
Choose and check a scaling approach for the task:
- Inspect units, distributions and missing values in numeric features.
- Select a transformation appropriate to the model and the data.
- Fit scaling parameters on training data and apply them to evaluation data.
- Compare alternatives on the same evaluation records and task metric.
How is Feature Scaling different from data normalization?
- Feature scaling is a broad term for changing the scale of numeric inputs.
- Data normalization can mean rescaling values in statistics and machine learning, or organizing database tables to reduce redundancy. Context determines the meaning.
- Standardization and min-max scaling are specific transformations; a team should name the method and parameters when documenting an experiment.
Where does Feature Scaling break in AI work?
Fitting a scaler before separating training and test records introduces data leakage. The scikit-learn guidance on common pitfalls recommends splitting before preprocessing and excluding test data when fitting transformations.
Extreme values can compress the bulk of a feature into a narrow part of a min-max range. Review outliers and compare suitable transformations on the same evaluation records. Decision tree models usually need little scaling, although the full modeling pipeline may include other scale-sensitive steps.