A few extreme values can strongly influence some scaling procedures. Decide whether those values are genuine observations, entry errors or a separate population before transforming them.
Scikit-learn documents multiple preprocessing scalers with different responses to distributions and outliers. Scikit-learn — preprocessing.
Review the unusual records
For a dimension dataset, check units and decimal placement before adopting a robust transformation. A preprocessing choice should not hide a correctable data error.
Evaluate the downstream effect
Compare suitable approaches inside the training and evaluation workflow. Keep physical units available in the source record even when the model uses transformed values. A scaled coordinate is an analytical representation, not a new product measurement specification.
Source reviewed 25 October 2026. Interpretation and original diagrams: PLUSHIP / Qunhe. Examples are illustrative.
Related reading
Combining Models Helps Only Under the Actual Evaluation Conditions
Channel Order Can Turn a Correct Image Into the Wrong Model Input
Object Detection Can Locate a Detail Without Measuring Its Quality