Hands-On Exercise: Model Explainability CT-AI Practice Questions (Page 4)

Questions 16–20

1. 16application · medium

   A bank's fraud-detection model uses a gradient-boosted tree ensemble with 500 features. A regulator asks for a per-transaction explanation of why a specific loan application was declined. The ML engineer needs a method that produces locally faithful explanations without retraining the model. Which approach should the engineer choose?

   • AA decision tree trained on all historical transactions, because it is inherently interpretable.
   • BGlobal feature importance from the tree ensemble, because it shows which features matter most overall.
   • CPartial dependence plots, because they show the average effect of a feature on predictions.
   • DLIME, because it fits an interpretable surrogate model around a single prediction.
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   A team is testing a high-stakes fraud-detection model. They need to explain a single suspicious transaction to a compliance officer. The model is a complex ensemble, and the team has access to both LIME and SHAP. The compliance officer needs an explanation that is both locally accurate and consistent with the model's behavior across similar transactions. Which method should the team prioritize, and why?

   • ALIME, because it is model-agnostic and can be applied to any ensemble.
   • BLIME, because it is faster to compute and easier to explain to a non-technical audience.
   • CSHAP, because it provides consistent and locally accurate attributions that satisfy the need for a reliable explanation.
   • DSHAP, because it always provides the true causal effect of each feature.
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A team is testing a random-forest model for loan approval. They want a global, model-agnostic explanation of which features are most important. The team needs a method that does not depend on the internal structure of the model and can be applied to any classifier. Which technique should they use?

   • ASHAP, because it provides local explanations for each prediction.
   • BGini importance from the random forest, because it is computed directly from the tree splits.
   • CLIME, because it fits a local surrogate model around each prediction.
   • DPermutation feature importance, because it measures the drop in model performance when a feature's values are randomly shuffled.
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   A healthcare startup uses a deep neural network to predict patient readmission risk. The clinical team wants to know which input features (e.g., age, lab values, prior admissions) most influenced each prediction. The model is a black box, and the team needs a method that provides consistent, mathematically grounded attributions across all features. Which explainability method should the team apply?

   • ASHAP, because it computes additive feature attributions based on cooperative game theory.
   • BLIME, because it is model-agnostic and fast to compute for any black box.
   • CA single decision tree surrogate, because it provides a visual and interpretable approximation of the network.
   • DPermutation feature importance, because it directly measures the drop in model performance when a feature is shuffled.
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   When evaluating the quality of an explanation from LIME, which aspect is most important to check?

   • AThe explanation is consistent with the model's actual behavior in the local region.
   • BThe explanation is identical for all instances in the dataset.
   • CThe explanation is generated in under one second.
   • DThe explanation uses the fewest possible features.
   Reveal Solution Discussion Select an answer first