Breaking Bias: Alpha Weighted Loss in Multi-objective Learning Taming Gender Stereotypes

Md Nur Amin, Abdullah Al Imran, Fatih S. Bayram, Lea Hübner, Alexander Jesser

Abstract: Navigating the uncertainties of job classification and gender bias, this paper presents multi-objective learning approach using BERT-based model that concurrently handles maximizing accuracy and mitigating gender bias. Main contribution of this study is making use of a loss function with a trade-off parameter, acknowledging no definitive ‘optimal’ solution is presumed. Eliminate unwanted bias or refrain systems from reinforcing bias seeking to unjust impact on people to sensitive characteristics is a critical consideration. This research underscores the pivotal role of decision-making under uncertainty in AI, setting a precedent for more conscious, bias-aware AI system design.

External IDs:doi:10.1007/978-981-99-8346-9_20