Hands-On Exercise: Model Explainability CT-AI Practice Questions (Page 5)

Questions 21–25

1. 21expert · hard

   A team is testing a deep learning model for medical image analysis. They use LIME to explain a prediction, but the explanations are unstable and vary significantly between runs. The team needs a more reliable explanation method. They have access to SHAP, but it is computationally expensive for the large image dataset. What is the most appropriate trade-off to consider?

   • AContinue using LIME because the instability is acceptable for image data.
   • BUse a simpler model, such as logistic regression, to avoid the need for complex explanations.
   • CUse SHAP on a representative subset of images to balance the need for reliable explanations with computational cost.
   • DUse SHAP on all images, regardless of the computational cost, to ensure maximum reliability.
   Reveal Solution Discussion Select an answer first
2. 22foundation · easy

   A LIME explanation for a text classification model shows that the word 'excellent' has a positive weight and the word 'poor' has a negative weight. What is the most reasonable interpretation?

   • AThe model is biased against the word 'poor'.
   • BThe words are the most frequent in the training data.
   • CThe words are the most important globally across all predictions.
   • DThe presence of 'excellent' increases the likelihood of the predicted class, while 'poor' decreases it.
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   A team is comparing explainability techniques for a tabular dataset. They need a method that provides a global ranking of feature importance and also allows them to understand the direction of the relationship (positive or negative) between each feature and the prediction. Which technique best satisfies both requirements?

   • AA partial dependence plot, because it shows the average effect of a feature on the prediction.
   • BSHAP summary plot, because it shows both the global importance ranking and the direction of each feature's impact.
   • CLIME, because it provides local explanations for individual predictions.
   • DPermutation feature importance, because it provides a clear ranking of features.
   Reveal Solution Discussion Select an answer first
5. 24application · medium

   A QA team is testing an AI system that recommends personalized content. They use LIME to generate explanations for the recommendations. The team wants to verify that the explanations are faithful to the model's actual behavior. What is the most appropriate way to assess the fidelity of the LIME explanations?

   • AAsk a domain expert to review the explanations for plausibility.
   • BCompare the surrogate model's predictions on the perturbed samples with the original model's predictions on the same samples.
   • CEnsure that the explanations include all features used by the original model.
   • DCheck that the explanations are consistent across multiple runs on the same input.
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   In the context of model explainability, what does 'feature importance' typically refer to?

   • AThe order in which features are processed by the model during training.
   • BThe number of times a feature appears in the training data.
   • CA measure of how much each feature contributes to the model's predictions.
   • DThe accuracy of the model when a feature is removed.
   Reveal Solution Discussion Select an answer first