- Block 13 — Class imbalance
- Objective
- define class imbalance rigorously and measure it; separate
imbalance as a property of a dataset from imbalance as a problem for learning,
and identify the three conditions that actually cause harm — class overlap, an
insufficient absolute count of minority examples, and small-disjunct structure;
establish the degrees of imbalance and the failure mode proper to each;
demonstrate the vacuity of accuracy on an imbalanced set; determine what each
metric measures when the classes are unequal; identify the four mechanisms by
which standard algorithms fail; trace imbalance back to its source and decide
what that source permits; and state the problem in the terms that govern it,
which are the asymmetric costs of the two error types.
- Estimated duration
- 75 minutes
- Prerequisites
- chapters 026, 027, 029, 030, 034, 035 and 049, and chapter
011 for the multiclass and multilabel vocabulary
- Associated quizzes
050.1-quiz-definition-and-measurement.md to
050.7-quiz-diagnostic-protocol.md