Answer: Measure class imbalance on the training dataset. Adapt the training process accordingly.

## Explanation To develop an unbiased machine learning model for loan allocation decisions across different demographic groups, the bank must address potential biases in the training data and model development process. Among the given options: **Option D (Correct):** "Measure class imbalance on the training dataset. Adapt the training process accordingly." This is the optimal approach because: - **Class imbalance** occurs when certain demographic groups are underrepresented in the training data relative to others. For example, if historical loan data contains significantly more approvals for one demographic group than another, the model may learn to favor the majority group, leading to biased predictions against minority groups. - **Measuring class imbalance** involves analyzing the distribution of data across demographic groups to identify disparities. - **Adapting the training process** includes techniques such as: - **Resampling methods** (oversampling minority classes or undersampling majority classes). - **Algorithmic adjustments** like using class weights during training to penalize errors on minority groups more heavily. - **Fairness-aware algorithms** that explicitly optimize for equity metrics. - This approach directly tackles data bias, which is a root cause of model bias, aligning with AWS best practices for responsible AI and fairness in ML. **Why other options are less suitable:** - **Option A:** "Reduce the size of the training dataset." This is counterproductive, as smaller datasets can exacerbate bias by reducing representation of minority groups and increasing variance in model performance. - **Option B:** "Ensure that the ML model predictions are consistent with historical results." This is incorrect because historical loan data often reflects existing societal or institutional biases (e.g., discriminatory lending practices). Enforcing consistency would perpetuate these biases rather than mitigate them. - **Option C:** "Create a different ML model for each demographic group." This is impractical and inefficient. It could lead to overfitting, increased complexity, and maintenance challenges, without necessarily addressing underlying data imbalances or ensuring fairness across groups. In summary, Option D is the most effective as it proactively identifies and mitigates bias at the data level, which is foundational to developing unbiased ML models in sensitive applications like loan allocation.

Author: LeetQuiz Editorial Team