Improving Fairness of AI Systems with Lossless De-biasing

Bias in AI decision-making is widespread and consequential: The paper identifies that AI systems used for critical decisions like credit scoring and patient triage frequently exhibit bias against underrepresented groups, making fairness improvements an urgent priority.

Traditional de-biasing approaches may unnecessarily sacrifice accuracy: Existing methods that eliminate demographic information from data may create an artificial trade-off between fairness and accuracy. This paper demonstrates that fairness improvements don't always require compromising system performance.

Oversampling underrepresented groups can improve both fairness and accuracy: The key finding is that increasing the representation of disadvantaged groups in training data simultaneously reduces algorithmic bias AND improves overall accuracy by addressing class imbalance—suggesting these goals can be complementary rather than competing.

Information-preserving techniques outperform information-removing ones: Unlike approaches that strip out demographic data, this "lossless" method retains demographic information while strategically rebalancing it, achieving better results on multiple fairness metrics across real-world datasets.

Fairness improvements require addressing data scarcity, not just removing information: The paper reframes the bias problem as fundamentally rooted in insufficient data for disadvantaged groups rather than the presence of demographic information itself, suggesting different solutions than previous approaches.