Explainable AI and SHAP Quiz

Answers and explanations appear as you take the quiz.

1. What mathematical concept are SHAP values based on?
   • The Shapley value from cooperative game theory
   • Bayesian posteriors
   • Gradient descent
   • Principal component analysis
2. The 'additivity' (local accuracy) property of SHAP means that for a single prediction, the SHAP values...
   • Are all positive
   • Sum to the difference between the prediction and a baseline output
   • Are normalized to lie between 0 and 1
   • Equal the model's training accuracy
3. Why is computing exact Shapley values generally expensive?
   • It requires retraining the model many times
   • It needs labeled test data for every prediction
   • It must average a feature's marginal contribution over all possible subsets of features
   • It only works on GPUs
4. Which SHAP variant computes exact values efficiently for gradient-boosted tree models?
   • KernelSHAP
   • DeepSHAP
   • LinearSHAP
   • TreeSHAP
5. Why do teams use explanation methods such as SHAP during model development?
   • To inspect model behavior, investigate influential features, and support review
   • Because SHAP proves the model is fair
   • Because every law requires SHAP for every model
   • Because feature attributions establish causal effects
6. What can a SHAP summary plot show across a dataset?
   • Whether a feature caused the outcome in the real world
   • How much features affect model outputs across examples, including the spread of their SHAP values
   • Which individual training row leaked the target label
   • A guaranteed ranking of features for every possible dataset
7. What does a local SHAP explanation describe?
   • A causal effect of changing a real-world feature
   • A guarantee that the model is fair
   • Feature attributions for one prediction relative to a chosen baseline and feature convention
   • The global training accuracy for every dataset
8. Why can correlated features complicate SHAP interpretation?
   • Correlated features make all SHAP values exactly zero
   • Correlation proves one feature caused the prediction
   • SHAP does not use a model output
   • Different assumptions about feature dependence can change how contributions are allocated