Google Professional Machine Learning Engineer Question 496
Single answerGoogle Cloud PlatformYou are working as a Machine Learning Engineer for a financial institution that uses a credit scoring model to predict the likelihood of loan defaults. The compliance team has raised concerns about the lack of transparency in the model's predictions and demands explanations for individual decisions. Which approach should you use to address this requirement, ensuring compliance while maintaining model performance?
- A
Use SHAP (SHapley Additive exPlanations) to generate feature importance values for individual predictions.
- B
Retrain the model using AutoML Tables Explainable Predictions to inherently provide feature attributions.
- C
Deploy the model with Vertex Explainable AI and enable feature attribution explanations.
- D
Replace the existing model with a simple linear regression model for better interpretability.
Show answer and explanation
Correct answer: C
Explanation
Vertex Explainable AI provides built-in tools for feature attribution and interpretability, allowing you to explain individual predictions without needing to retrain or replace the existing model. This enables compliance with transparency requirements while maintaining model performance.
- A. Incorrect.
SHAP is a powerful explainability tool, but it is not natively integrated with Google Cloud services. While it could work, it does not leverage the native capabilities of Vertex AI for explainable AI.
- B. Incorrect.
AutoML Tables supports explainable predictions, but this would require retraining the model, which may not be feasible if the current model is already optimized for performance.
- C. Correct.
Deploying the model with Vertex Explainable AI allows you to use integrated feature attribution explanations without retraining the model, meeting the compliance team's demands while maintaining the existing model's performance.
- D. Incorrect.
Replacing the model with linear regression sacrifices model performance unnecessarily. Explainability can be achieved without compromising the predictive power of the current model.