AI-900 Practice Question: Describe fundamental principles of machine learning on Azure

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Explaining the model's predictions using a simplified version of the model that is easier to interpret

  Why it's wrong here

  SHAP values explain individual predictions by quantifying each feature’s contribution using game-theoretic Shapley values, not by substituting a simplified model. This option describes a surrogate model approach (e.g., a linear approximation of a black-box model), which is a different technique for interpretability. It is tempting because surrogate models also aim to make predictions understandable, and in scenarios where global, model-agnostic explanations are sufficient—rather than per-instance feature attribution—a simplified model would be correct.

• ✓

  Calculating each feature's contribution to a specific prediction to explain why the model made that decision

  Why this is correct

  SHAP (SHapley Additive exPlanations) values assign each feature a numeric contribution to a model's output for one specific instance, derived from cooperative game theory's Shapley values. They satisfy desirable properties such as local accuracy (the sum of attributions equals the prediction), consistency (changing a model so a feature has greater impact never lowers its attribution), and missingness (features with no influence get zero attribution). This is precisely a local, per-prediction explanation: it quantifies how much each input feature pushed the prediction away from the baseline toward the final decision.

• ✗

  Displaying the model's source code so users can verify what computations are performed

  Why it's wrong here

  Displaying the model's source code reveals the exact computational steps and hyperparameters implemented, which is a transparency measure often used in open-source or audit settings. However, source-code visibility is not interpretability of the learned model's behavior: even with full code access, you still cannot determine how each feature influenced a specific prediction without executing the model and analyzing its internals. SHAP, by contrast, is a model-agnostic post-hoc explanation technique that works on the trained model's outputs (or predictions) and does not require modifying or viewing the underlying code. The objective is to explain the reasoning behind a prediction, not to disclose implementation details.

• ✗

  Testing the model on a separate evaluation dataset to report overall accuracy

  Why it's wrong here

  Evaluating a model on a separate holdout or test dataset measures aggregate predictive performance (e.g., accuracy, precision, recall, F1 score) across many examples. While this is a crucial step for estimating generalization error, it reports only how often the model is correct overall and says nothing about which features drove any individual prediction. In contrast, SHAP provides instance-level feature attributions for a single decision, answering "why this particular outcome" rather than "how well does the model perform in general." These are complementary but conceptually distinct tasks.

Go deeper

Related to this question