- Block 6 — Transforming the data
- Objective
- state precisely why a smaller feature set can outperform a
larger one, define the three families of selection methods — filter, wrapper,
embedded — with the assumptions each one carries, work the univariate tests by
hand and know exactly what each of them can and cannot see, control the false
discovery rate when thousands of features are screened at once, measure the
combinatorial cost of wrappers, read an importance measure without being
deceived by cardinality bias or by correlated columns, assess the stability of
a selection across resamples, and apply without exception the rule that a
selection fitted on the target belongs inside the cross-validation loop.
- Estimated duration
- 60 minutes
- Prerequisites
- chapters 016 (correlation and causation), 017
(multicollinearity), 023 (feature scaling) and 024 (feature engineering)
- Associated quizzes
025.1-quiz-why-select.md to
025.8-quiz-choosing-a-method.md