- Block 13 — Class imbalance
- Objective
- turn the diagnosis of chapter 050 into a procedure ordered by
cost and risk; derive the Bayes-optimal threshold and establish by measurement that
moving it is the first remedy and usually the sufficient one; show that class weighting
is a threshold move plus a rotation of the boundary; define random oversampling, random
undersampling, SMOTE and its variants, the informed undersamplers and the combined
methods precisely enough to reproduce them, and measure what each is worth; establish
the rule that resampling belongs inside the cross-validation fold and quantify the
fiction produced by breaking it; report honestly where synthetic oversampling helps and
where it hurts; frame the subject as expected-cost minimization; cover the
imbalance-aware ensembles, the one-class framing, the return on collecting positives,
and the calibration damage every remedy causes; and close with a decision table.
- Estimated duration
- 95 minutes
- Prerequisites
- chapter 050 in full; chapters 026, 027, 028, 034 and 035 for
the protocol; chapter 049 for the model families; chapter 030 for the distinction
between a loss and a metric
- Associated quizzes
051.1-quiz-the-remedy-ladder.md to
051.8-quiz-decision-procedure.md