Overview of Bias Mitigation Strategies
What You'll Learn
You'll understand the three main stages where bias can be addressed in ML systems: before training, during training, and after training.
The Three-Stage Mitigation Framework
Now that you know how to measure bias, how do you reduce it? Bias mitigation strategies fall into three categories based on when they're applied in your ML pipeline:
Pre-processing Approaches
These techniques modify your training data before model training begins. Think of this as cleaning your ingredients before cooking. Strategies include:
- Reweighting examples to balance representation across protected groups
- Sampling techniques (oversampling underrepresented groups, undersampling overrepresented ones)
- Learning fair representations that remove discriminatory information while preserving predictive features
- Relabeling or repairing biased labels in the dataset
In-processing Approaches
These modify the training algorithm itself to optimize for both accuracy and fairness simultaneously. Imagine adding fairness constraints while teaching the model. Examples include:
- Adding fairness constraints to the loss function (e.g., penalizing demographic parity violations)
- Adversarial debiasing (training a model to be accurate but fooling an adversary trying to predict protected attributes)
- Regularization terms that enforce fairness metrics
Post-processing Approaches
These adjust model predictions after training, without retraining. Think of this as adjusting seasonings after cooking. Techniques include:
- Threshold optimization (using different decision thresholds for different groups to achieve equalized odds)
- Calibration adjustments to ensure predictive parity
- Reject option classification (withholding predictions in uncertain regions near decision boundaries)
Each approach has trade-offs: pre-processing is model-agnostic but may lose information; in-processing is integrated but model-specific; post-processing is flexible but may sacrifice overall accuracy.
Key Takeaway: Bias mitigation can happen at three stages—pre-processing (data), in-processing (algorithm), and post-processing (predictions)—each with distinct trade-offs between fairness, accuracy, and implementation complexity.