Google Professional Machine Learning Engineer Question 497
Single answerGoogle Cloud PlatformYou are deploying a machine learning model for credit risk assessment in a bank. The stakeholders require that the model's predictions are interpretable so that loan officers can explain to customers why a loan application was approved or denied. Which approach would be most suitable to achieve explainability in this scenario?
- A
Use SHAP (SHapley Additive exPlanations) values to provide feature-level contributions for each prediction.
- B
Deploy the model with no additional tools and rely on its high accuracy during testing.
- C
Use a deep learning model with many layers to ensure the model captures the most complex patterns in the data.
- D
Train a simpler linear model instead, as it is inherently interpretable and does not require further explainability tools.
Show answer and explanation
Correct answer: A
Explanation
SHAP values are a standard method for achieving explainability in machine learning models. They can provide insights into how each feature contributes to a specific prediction, addressing the stakeholders' requirement for interpretability. This approach allows loan officers to explain decisions to customers while maintaining the performance of the deployed model.
- A. Correct.
SHAP values are widely used for explaining individual predictions by showing the contribution of each feature to the model's output, making it highly suitable for explaining decisions in a credit risk model.
- B. Incorrect.
While high accuracy is important, it does not address the need for interpretability or explainability that stakeholders require in this scenario.
- C. Incorrect.
Deep learning models are often considered black boxes and are not inherently interpretable, making them unsuitable for this scenario where stakeholders need clear explanations.
- D. Incorrect.
Linear models are interpretable but may sacrifice performance if the problem requires capturing complex patterns in the data. The requirement is to explain predictions of the current model, not to replace it entirely.