Ch 5: Remove Bias from ML Model¶

Introduction¶

This chapter covers model-agnostic techniques to overcome data bias during or before model training. These approaches work with most ML algorithms and help make the final model more explainable.

Technique	When Applied	Handles	Algorithm Support
Reweighting	Before training	One protected feature at a time	Classification only
Additive Counterfactual Fairness	During training	Multiple protected features	Classification & regression

Key Insight¶

Don't Change the Data — Change the Weights

Reweighting achieves bias reduction without altering any data points or labels. It adds sample weights that the loss function uses during optimization. This preserves data integrity while reducing discrimination.

Choosing Your Approach¶

graph TD
    A[How many protected features?] --> B{One?}
    B -->|Yes| C{Classification?}
    C -->|Yes| D[✅ Reweighting]
    C -->|No| E[Use ACF]
    B -->|Multiple| F[Create Composite Feature<br>or Use ACF]
    F --> G{Can combine features?}
    G -->|Yes| D
    G -->|No| E[✅ Additive Counterfactual Fairness]

What to Expect¶

Reweighting: Little to no accuracy loss while significantly improving fairness metrics
ACF: Handles complex multi-feature bias but requires additional model training for residual computation

Off-the-Shelf Tools

IBM's AIF360 and Microsoft's Fairlearn provide packaged implementations, but this chapter covers the underlying techniques so you can understand, customize, and optimize them for your specific problem.

Next: Reweighting →