Utrata Hubera i solidny quiz dotyczący regresji

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. How does Huber loss treat a residual far beyond its threshold?
   • Its penalty grows linearly with residual magnitude
   • Its penalty becomes exactly zero
   • It squares the residual at a faster rate
   • It removes the observation automatically
2. What behavior does Huber loss use for small residuals?
   • A constant penalty
   • A quadratic penalty
   • A binary outlier label
   • No contribution to the objective
3. How does RANSAC differ from Huber regression in the described approach?
   • It estimates only class probabilities
   • It assigns the same squared penalty to every point
   • It seeks a consensus set from repeatedly fitted subsets
   • It removes the need to choose any threshold
4. A large residual is a valid, important rare event. What should the analyst consider before using Huber loss?
   • Whether rarity alone is enough to treat the event as a measurement error
   • Whether Huber loss will automatically relabel the observation as an inlier
   • Whether all rare observations should be removed before evaluation
   • Whether downweighting that event conflicts with the decision goal
5. Why can an extreme residual occur even when the record is not erroneous?
   • Every large residual is a data-entry mistake
   • It may be a valid rare event or reflect model misspecification
   • Huber loss creates all large residuals
   • A linear predictor cannot have residuals
6. When should a robust threshold be selected?
   • Tune repeatedly on final-test errors because they best represent deployment
   • Choose the threshold that minimizes training loss, then report that same training score
   • Select it using training or validation data while reserving the final test set
   • Use the library default without checking whether it suits the residual scale
7. Which feature can remain a concern for robust regression despite reduced sensitivity to large response residuals?
   • Only whether the average response residual is zero
   • Only whether the residual histogram has a symmetric center
   • Only whether the response variable was standardized
   • Whether unusual predictor-space leverage can still affect the fit
8. Why report tail performance when extreme errors matter operationally?
   • A robust objective may deemphasize cases that dominate operational tail risk
   • Huber loss assigns every extreme observation exactly zero influence
   • Training loss alone measures accuracy on unseen extreme outcomes
   • The average residual identifies the size of every tail error