Prueba de AdaBoost

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

1. How are AdaBoost's weak learners typically trained relative to one another?
   • Independently under unchanged sample weights
   • All at once using one common unweighted loss
   • Sequentially, with later rounds adapting to earlier errors
   • By retaining only the final learner and discarding prior rounds
2. After a learner misclassifies an example, what usually happens to its relative training weight?
   • It becomes identical to the learner's ensemble vote
   • It increases for a subsequent round
   • It is permanently removed after its first misclassification
   • It is reset to equal weight before every subsequent round
3. In the classic binary formulation, what kind of weighted error earns a useful positive learner vote?
   • An error greater than one-half
   • Any error because the ensemble weight is always positive
   • An error measured only on the final test set
   • A weighted error below one-half
4. Why can mislabeled outliers be problematic for AdaBoost?
   • They may keep receiving increased weight
   • Reweighting proves that their labels must be correct
   • Each stump automatically removes examples with inconsistent labels
   • Repeated rounds guarantee such examples have no influence
5. A one-split decision tree is known by what common nickname?
   • Decision forest
   • Decision branch
   • Decision stump
   • Decision ensemble
6. What combines the sequence of learners into the final AdaBoost prediction?
   • Only the last learner's prediction
   • A weighted vote of their outputs
   • A vote with necessarily equal weights for every learner
   • The prediction of the learner trained on the largest subset alone
7. What should a practitioner inspect when several examples remain difficult across rounds?
   • Whether high weight alone proves every case is mislabeled
   • Whether all difficult cases should automatically be removed
   • Whether increasing rounds guarantees that all cases become representative
   • Whether they are valid, mislabeled or outside intended use
8. Which statement about more boosting rounds is justified?
   • More rounds do not guarantee better generalization
   • Lower training error guarantees lower future error
   • Repeated weighting guarantees correction of noisy labels
   • A large ensemble eliminates the need for held-out evaluation