Towards Understanding Fairness and its Composition in Ensemble Machine Learning

Fair individual models don't guarantee fair ensembles: The research demonstrates that even when individual machine learning models are fair, combining them into an ensemble can produce unfair results, revealing that fairness "composes in a non-trivial way" across different ensemble types (bagging, boosting, stacking, and voting).

Ensemble design choices matter for fairness: The study found that hyperparameters—the settings that control how multiple models are combined—significantly influence the overall fairness of ensemble systems, suggesting that fairness can be improved through thoughtful design choices rather than relying solely on bias mitigation techniques.

Real-world ensembles show fairness problems: Analysis of 168 ensemble models from Kaggle across four fairness datasets revealed documented fairness implications for minority groups based on protected characteristics like race, sex, and age, indicating this is a practical problem affecting deployed systems.

Different ensemble methods behave differently: The comprehensive study of popular ensemble approaches (bagging, boosting, stacking, and voting) suggests that fairness composition varies by algorithm type, meaning practitioners need algorithm-specific understanding rather than one-size-fits-all solutions.

Fairness can be improved without additional mitigation: The findings indicate that ensembles can be designed to be fairer through better configuration of existing methods, offering a potentially simpler path to fairness than implementing separate bias-reduction techniques.