A large retail bank wants an ML system to guide loan allocation decisions across different demographic groups. What action is needed to help ensure the model is unbiased?

Assessing class imbalance and adjusting the training process is crucial for mitigating bias. If certain demographic groups are underrepresented in the training data (class imbalance), the model may not learn to make accurate or fair predictions for those groups, leading to biased outcomes. Techniques like oversampling, undersampling, or using weighted loss functions can address this. Reducing the training dataset size is counterproductive as it often leads to less robust and more biased models. Forcing predictions to match historical decisions can perpetuate existing human biases present in those decisions. Building separate models for each demographic group might be overly complex and could still lead to bias if individual group datasets are imbalanced or if the models are not evaluated for fairness across groups.