Addressing bias in bagging and boosting regression models.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Growing use of AI necessitates bias investigation in ML models.
• Prior research focused on classification, neglecting regression model bias.
• Regression models, crucial for many applications, often exhibit performance disparities.

Purpose of the Study:

• To present a novel, accessible methodology for mitigating bias in regression models.
• To extend bias mitigation techniques to bagging and boosting ensemble methods.
• To rigorously measure and reduce bias related to protected attributes.

Main Methods:

• Developed a bias mitigation technique applicable to models minimizing differentiable loss functions.
• Integrated a regularization term into the model's loss function to penalize error correlations with protected attributes.
• Applied and validated the methodology on Random Forest, Gradient-Boosted Trees, and XGBoost models.

Main Results:

• The proposed method effectively reduces bias in tree-based ensemble regression models.
• Bias in models predicting road-level traffic volume was reduced by over 50% for minority-populated areas.
• Despite high overall accuracy, baseline models showed poor performance on roads in minority areas.

Conclusions:

• The presented methodology offers an effective solution for mitigating bias in regression ML models.
• This approach enhances fairness and equitable performance across different demographic groups.
• The technique is broadly applicable to various ML models trained via differentiable loss minimization.