Google Professional Machine Learning Engineer Question 476
Select 3Google Cloud PlatformYou are designing an AI solution to predict loan approvals for a financial institution. During testing, you discover that the model systematically denies loans to applicants from a specific demographic group, even though the decision is not justified by any input features. Which steps should you take to assess and address fairness and bias in the solution?
- A
Perform a feature importance analysis to check if the model is relying heavily on sensitive attributes such as race, gender, or age.
- B
Use a fairness metric, such as disparate impact or equalized odds, to evaluate bias in the model's predictions.
- C
Exclude all sensitive attributes from the training data to ensure the model does not discriminate.
- D
Retrain the model using a balanced dataset that ensures equal representation of all demographic groups.
- E
Perform a differential analysis to assess disparities in model performance across different demographic groups.
Show answer and explanation
Correct answers: A, B, E
Explanation
Assessing AI solution readiness for fairness and bias involves identifying issues in the model's predictions and performance. Feature importance analysis and fairness metrics like disparate impact help detect and measure bias, while a differential analysis ensures the model performs equitably across different groups. Simply removing sensitive attributes or balancing the dataset are insufficient alone, as bias can still emerge through indirect correlations or other factors.
- A. Correct.
Correct: Feature importance analysis can help identify if the model is overly dependent on sensitive attributes, which may lead to biased predictions.
- B. Correct.
Correct: Fairness metrics like disparate impact or equalized odds are specifically designed to measure bias in model predictions and are essential for assessing fairness.
- C. Incorrect.
Incorrect: Simply excluding sensitive attributes from the training data does not guarantee the absence of bias, as the model might infer protected attributes from correlated features.
- D. Incorrect.
Incorrect: While balancing the dataset can help with fairness, it is not sufficient by itself. Further evaluations and fairness metrics are needed.
- E. Correct.
Correct: Differential analysis helps identify disparities in model performance (e.g., accuracy, precision) across demographic groups, which is a key step in assessing bias.