FairPO: Robust Preference Optimization for Fair Multi-Label Learning
Soumen Kumar Mondal, Prateek Chanda, Akshit Varmora, Ganesh Ramakrishnan
TL;DR
FairPO addresses fairness in multi-label classification by partitioning labels into a privileged subset to be aggressively improved and a non-privileged subset to be safeguarded. It introduces a targeted preference-based objective for privileged labels and a constrained objective for non-privileged labels, balanced via Group Robust Preference Optimization. The framework supports reference-based (DPO) and reference-free (CPO, SimPO) variants and demonstrates significant gains on tail labels with minimal impact on head labels in MS-COCO and NUS-WIDE. Ablation studies validate the necessity of each component, and the approach offers a practical path to robust, fair MLC in real-world settings.
Abstract
Multi-label classification (MLC) often suffers from performance disparities across labels. We propose \textbf{FairPO}, a framework combining preference-based loss and group-robust optimization to improve fairness by targeting underperforming labels. FairPO partitions labels into a \textit{privileged} set for targeted improvement and a \textit{non-privileged} set to maintain baseline performance. For privileged labels, a DPO-inspired preference loss addresses hard examples by correcting ranking errors between true labels and their confusing counterparts. A constrained objective maintains performance for non-privileged labels, while a Group Robust Preference Optimization (GRPO) formulation adaptively balances both objectives to mitigate bias. We also demonstrate FairPO's versatility with reference-free variants using Contrastive (CPO) and Simple (SimPO) Preference Optimization.