Partial Dependence Plots Quiz

Answers and explanations appear as you take the quiz.

1. How does a single-feature partial dependence value at one grid point arise?
   • Average observed feature values without model predictions
   • Fit a new model using only the selected column
   • Average model predictions after setting that feature to the grid value
   • Count how many examples have that exact value
2. What does an ICE display add beyond a partial dependence average?
   • A causal effect estimate
   • A prediction curve for each individual example
   • A guarantee that features are independent
   • A new model trained for every row
3. ICE lines fan apart as a feature changes. What does this suggest?
   • The average represents every example's response equally well
   • The selected feature has the same additive response for every case
   • Only the baseline prediction levels differ while curve shapes remain parallel
   • The average may hide heterogeneous model responses
4. Why can correlated predictors make a partial dependence plot misleading?
   • Substitution can create unrealistic feature combinations
   • Averaging automatically removes all feature dependence
   • Keeping other inputs fixed guarantees a valid intervention
   • Restricting a feature to its observed range guarantees every joint combination is realistic
5. What kind of claim does a partial dependence curve support most directly?
   • The causal effect of changing a feature
   • The model's calibration on every subgroup
   • How this fitted model's average output varies under the plot calculation
   • The distribution of true labels at each feature value
6. For a classifier, what should a report specify about the plotted output?
   • Only whether a classifier exceeds 50% accuracy
   • Class and output scale, such as probability or decision score
   • Only the ranking of feature variances
   • Only whether the predictions were rounded to labels
7. Which alternative may reduce extrapolation concerns under dependent features by estimating local changes?
   • A marginal plot without checking confounding inputs
   • An uncentered ICE plot that uses the same unrestricted substitutions
   • A global feature ranking based only on training coefficients
   • Accumulated local effects
8. Centered ICE is useful for comparing what across examples?
   • Differences in curve shape after subtracting a reference prediction
   • Only differences in the original prediction baselines
   • The distribution of observed outcomes without model predictions
   • Causal treatment effects for each example