Explainability and Fairness in Machine Learning

Artificial Intelligence is finding its way to ever more applications. Nonetheless, it is increasingly required that decision-making procedures must be explainable and fair. As many applications are based on black-box models, there is a strong need for more explainable AI algorithms. For this reason, our paper explores the practical implications and effectiveness of four bias mitigation algorithms (learning fair representations, reweighing, Equality of Odds, and Reject Option based Classification) based on a standard XGBoost classifier to build an explainable and fair prediction model on a real-world loan dataset. The models were evaluated based on their performance, fairness, and explainability. Potential biases, i.e. fairness, were detected with the use of NLP techniques and evaluated with the AIF360 metrics, whereas the explainability of the model was tested by post-hoc explanations (SHAP method). The best results were obtained by the reweighing algorithm that improved the fairness while maintaining a high model performance and explainability.

Alexander Stevens, Peter Deruyck, Ziboud Van Veldhoven, Jan Vanthienen

Link to paper: here