Bias mitigation

Responsible AI Data Management

Maria Prokofieva

Lead ML engineer

We will cover

• Strategies throughout the project lifecycle
• Re-sampling and re-weighting
• Re-labeling and sensitivity attribute removal
• Adversarial training and model re-calibration
• User behavior
• Fairness constraints and algorithm selection
• Other strategies exist

Bias revisited

• Systematic errors or prejudices
• Unfair outcomes for groups or individuals

¹ Image by Streamline HQ

Bias across project

• In raw data
• In pre-processing
• In modeling
• In post-deployment

Bias in pre-processing

• Re-sampling
• Re-weighting
• Re-labeling
• Sensitivity attribute removal

Strategies in pre-processing

• Re-sampling:
  • Adjust distribution for balanced class representation
• Re-weighting:
  • Correct biases with class weights, keeping dataset size constant
• Re-labelling:
  • Update labels for accurate target representation
• Sensitivity attribute removal:
  • Eliminate protected attributes

Bias in modeling

• Bias from the algorithm
• Adversarial training
• Fairness constraints
• Algorithm selection with simpler, robust, or generalizable models

Financial advisor

• Bias related to regional economic disparities
• Adversarial training to correct bias
• Decision trees and regularized linear models

¹ Image by Streamline HQ

Bias in post-deployment and use

• Monitor bias to prevent model drift
• Model re-calibration
• Improper application use

¹ Image by Streamline HQ

Financial advisor

Model drift:

• Underrepresented freelance economy

Incorrect use:

• Updating financial information

¹ Image by Streamline HQ

Let's practice!

Responsible AI Data Management