In recent years, concerns about unfair decisions made by machine learning algorithms have increased, leading to the development of various fairness-aware methods, though most focus only on single-label classification. This paper addresses the gap in fairness for multi-label classification by extending Demographic Parity (DP) and Equalized Opportunity (EOp) to these scenarios. The authors propose a new framework, Similarity s-induced Fairness (sγ -SimFair), which improves fairness estimation by utilizing data with similar labels and demonstrates superior performance in both theoretical analysis and experimental results on real-world datasets.

[ Paper PDF ]