Hands-On Exercise: Model Explainability CT-AI Practice Questions (Page 3)

Questions 11–15

1. 11application · medium

   A data scientist uses LIME to explain a prediction from a logistic regression model. The LIME output shows that the feature 'number_of_open_accounts' has a negative weight in the local surrogate model. What does this indicate about the original model's prediction for this specific instance?

   • AThe model is biased against customers with many open accounts.
   • BThe feature contributed to a lower predicted probability for this instance.
   • CThe feature always decreases the model's predictions for all instances.
   • DThe feature is not important for the model's prediction.
   Reveal Solution Discussion Select an answer first
2. 12foundation · easy

   When applying SHAP to a trained model, what is the first required step?

   • ARemove all correlated features from the dataset.
   • BConvert the model to a neural network format.
   • CTrain an interpretable surrogate model on the entire dataset.
   • DSelect a background dataset to represent the typical input distribution.
   Reveal Solution Discussion Select an answer first
4. 13foundation · easy

   In the context of testing AI-specific quality characteristics, why is it important to evaluate the usefulness of explainability outputs?

   • ATo ensure the explanations are understandable and actionable for stakeholders.
   • BTo reduce the model's training time.
   • CTo increase the model's accuracy on the test set.
   • DTo replace the need for functional testing.
   Reveal Solution Discussion Select an answer first
5. 14expert · hard

   A team is testing a model that predicts insurance claim risk. They use SHAP to explain the model and find that the feature 'number_of_claims' has a high positive SHAP value for a specific customer. However, the team also notices that the model's prediction for this customer is low risk. What is the most appropriate conclusion?

   • AThe SHAP values are incorrect because a high number of claims should always increase risk.
   • BThe team should remove the number_of_claims feature because it is confusing.
   • CThe high SHAP value for number_of_claims is pushing the prediction higher, but other features are pushing it lower, resulting in a low-risk prediction.
   • DThe model is biased against customers with a high number of claims.
   Reveal Solution Discussion Select an answer first
6. 15application · medium

   A QA team is evaluating the quality of LIME explanations for a sentiment-analysis model. They notice that running LIME twice on the same review produces noticeably different feature attributions. Which quality characteristic of the explanations is most clearly compromised?

   • AStability, because the explanations are not reproducible for the same input.
   • BFidelity, because the surrogate model does not match the original model's predictions.
   • CSelectivity, because the explanation focuses on too few features.
   • DCompleteness, because not all features are included in the explanation.
   Reveal Solution Discussion Select an answer first