You recently developed a classification model that predicts which customers will be repeat customers. Before deploying the model, you perform post-training analysis on multiple data slices and discover that the model is under-predicting for users who are more than 60 years old. You want to remove age bias while maintaining similar offline performance. What should you do?

This approach compensates for bias directly in the data by enhancing the data distribution of users above 60 years old. Some useful preprocessing steps could be filling null values, bucketizing, clipping outliers, sampling, or even collecting new data. Removing correlated features leads to large drops in offline performance. Modifying input baselines only adjusts explainability, not model performance. Post-processing calibration is brittle, fixing the symptom rather than the cause, and can introduce other biases.