- Block 6 — Transforming the data
- Objective
- separate the two distinct mechanisms through which the scale of
a variable degrades a model — domination of the distance metric, and distortion
of the optimization surface; master standardization, min-max scaling and their
robust alternatives, including what each one bounds, what it does not bound and
how each behaves on values never seen during fitting; distinguish scaling from
the shape transformations that correct skewness; know for every family of
algorithms whether scaling is required, useful or strictly without effect, and
the technical reason why; and apply without exception the rule that scaling
parameters are estimated on the training set alone.
- Estimated duration
- 55 minutes
- Prerequisites
- chapters 006 (features and target), 014 (descriptive
statistics), 021 (outliers) and 022 (categorical encoding); the notion of a
train/test split (chapter 026)
- Associated quizzes
023.1-quiz-why-scale.md to
023.6-quiz-train-only-fitting-and-pipelines.md