Quiz dotyczący metod wyboru funkcji

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. How does feature selection differ from principal component analysis?
   • Selection and PCA both construct the same linear projections
   • Selection chooses data rows while PCA chooses original columns
   • Selection retains input variables; PCA constructs transformed components
   • Selection is necessarily unsupervised while PCA requires class labels
2. A univariate filter may miss a variable that is useful only through what?
   • Its variance considered in isolation
   • An interaction with another feature
   • The column's frequency of missing values alone
   • Its linear correlation with the target alone
3. What makes recursive feature elimination a wrapper method?
   • It computes only a label-free variance threshold
   • It transforms every feature into a principal component
   • It fixes an initial feature ranking without refitting
   • It repeatedly fits an estimator to judge feature subsets
4. An L1-regularized linear model sets some coefficients to zero during fitting. This is an example of which family?
   • Embedded selection
   • An estimator-independent univariate filter
   • An external search that refits every candidate subset
   • An unsupervised component projection
5. Why must supervised feature selection be repeated inside each training fold?
   • To guarantee identical selected variables in every fold
   • To avoid tuning the selected subset size
   • To prevent held-out labels from affecting the selected subset
   • To make training accuracy an unbiased test score
6. How should high mutual information with a target be interpreted?
   • Evidence of statistical dependence, not proof of causation
   • Proof the feature will help every estimator
   • Proof the feature is independent of others
   • A guarantee of stable performance in production
7. Two correlated variables split a tree model's importance. What should an analyst conclude?
   • Both variables are proven useless
   • The lower score proves negative causation
   • Importance scores are unaffected by correlation
   • Importance can be distributed among substitutes, so inspect redundancy
8. A selected subset changes greatly across resamples. Which diagnostic is most directly relevant?
   • Selection stability across resamples
   • The training score from a single fitted subset
   • The marginal variance of one selected column
   • The number of input variables before selection