Bias Detection in ML Models via Correlated Attribute Mapping

The system performs preliminary actions by training separate models to predict protected attributes from non-protected attributes before the main modeling process. These pre-trained models capture the correlations between protected and non-protected attributes, allowing the system to later adjust and correct biased predictions without retraining the entire model, thus maintaining accuracy while reducing bias.

The system introduces intermediary models that act as mediators between the input data and the final model output. These intermediary models predict protected attributes as intermediate steps, allowing the system to identify and correct biased patterns without directly removing the protective attributes from the training data, thereby preserving useful correlations while eliminating harmful bias.

The system segments the complex bias correction process into separate, manageable steps: first training models to predict protected attributes, then using these predictions to identify and correct biased data points, and finally integrating corrections into the main model. This segmentation reduces overall complexity by breaking down the data processing into distinct phases that can be optimized independently.

The system creates copies of the training data with corrected bias by mapping data points from the original dataset to adjusted representations. These copied datasets are then used for training the final model, allowing the system to preserve the structure and relationships of the original data while eliminating bias, thus reducing processing complexity compared to working with the original biased data directly.