Mitigating biases in AI systems trained on tabular data

Sammanfattning

As Artificial Intelligence (AI) becomes more common in decision-making systems, bias in AI, causing unfair consequences, becomes an important issue. This thesis aims to investigate and mitigate AI bias by utilizing the open-source fairness tools AI Fairness 360 and Fairlearn together, creating an ensemble system in the Python programming language. The system aims to investigate whether an ensemble approach with multiple AI predictors can reduce bias while maintaining accuracy and performance. The system is designed to be usable without coding skills and incorporates an automatic mode and a manual mode to provide value for users with different levels of field knowledge. It also presents performance evaluations of models to provide useful insight. The project results show that while bias cannot be entirely eliminated, the implemented ensemble approaches have some potential in removing parts of it. We also suggest improvements to the system, among them are explanations of key concepts, making the tool more informative, and adding more options to specialize model training to target specific biases.