One-Class SVM Quiz

Answers and explanations appear as you take the quiz.

1. What data does One-Class SVM commonly use to learn a novelty boundary?
   • A reference sample intended to represent normal behavior
   • Balanced labels for every anomaly and inlier
   • Only the future query points
   • A sequence of cluster IDs from k-means
2. What does a later point outside the learned boundary mean?
   • It is definitively a failure
   • It is novel under this fitted representation and warrants interpretation
   • It has a calibrated probability of being fraudulent
   • It must have been excluded from the reference data by error
3. For an RBF kernel, what does gamma control?
   • The exact fraction of future cases that are wrong
   • The number of anomaly classes
   • How quickly similarity declines with distance
   • The response variable's unit
4. How should nu be interpreted in the common formulation?
   • A lower bound on training errors and an upper bound on support vectors
   • An exact estimate of population anomaly prevalence
   • A probability that a future point is anomalous
   • An upper bound on training errors and lower bound on support vectors
5. Why scale features before fitting a kernel SVM?
   • Scaling sets the correct nu automatically
   • Feature units affect distances and therefore the kernel boundary
   • The kernel ignores all feature values after scaling
   • It creates labels for unknown anomalies
6. What can happen if reference training data contain many anomalous points?
   • The model becomes supervised by itself
   • Nu guarantees all anomalies are removed
   • The learned boundary may treat some anomalies as normal
   • The kernel changes to a Gaussian response family
7. Why can benign process drift create false novelty alerts?
   • All future points are used to refit the boundary
   • Gamma automatically detects drift labels
   • Nu is the exact real anomaly rate
   • A fixed boundary may no longer cover changed normal behavior
8. Which statement about the decision score is appropriate?
   • It reflects position relative to a fitted boundary, not a causal diagnosis
   • It is a calibrated probability by definition
   • It directly identifies why the point is unusual
   • It remains unchanged under every scaling choice