Mitigating implicit and explicit bias in structured data without sacrificing accuracy in pattern classification

• DOI: 10.1007/s00146-024-02003-0
• Publication Date: 2024-07-10T04:02:04Z
• Authors: Fabian Hoitsma; Gonzalo Nápoles; Çiçek Güven; Yamisleydi Salgueiro
• Keywords: Fair machine learning; 0202 electrical engineering, electronic engineering, information engineering; Bias mitigation; Instance reweighting

Abstract Using biased data to train Artificial Intelligence (AI) algorithms will lead decisions, discriminating against certain groups or individuals. Bias can be explicit (one several protected features directly influence the decisions) implicit indirectly decisions). Unsurprisingly, patterns are difficult detect and mitigate. This paper investigates extent which one more in structured classification sets mitigated simultaneously while retaining data’s discriminatory power. The main contribution of this concerns an optimization-based bias mitigation method that reweights training instances. algorithm operates with numerical nominal mitigate simultaneously. trade-off between accuracy loss controlled using parameters objective function. simulations real-world show a reduction up 77% complete removal at no cost wrapper classifier trained on data. Overall, proposed outperforms state-of-the-art methods for selected sets.