Title: Unveiling the Ideal Fair Classifier for Multi-Class Tasks

Abstract:

Achieving equitable treatment across various demographic groups remains a formidable hurdle in multi-class classification, largely because machine learning models often harbor deep-seated biases. While numerous strategies exist to mitigate bias, they are predominantly designed for binary classification contexts. Consequently, adapting these methods to multi-class environments is often impractical or ineffective, given the complexity of multi-dimensional outputs and sophisticated fairness metrics. This study addresses two critical, unresolved issues in fair classification: (i) defining the optimal trade-off boundary between accuracy and fairness in multi-class scenarios, and (ii) developing viable algorithms that can achieve this balance during various stages of the training process.

To address these issues, we introduce a probabilistic framework that allows for the analytical determination of the optimal classifier while adhering to fairness constraints. Leveraging this foundation, we present two attribute-blind algorithms designed to implement fairness in real-world applications. The first is an in-processing method that integrates fairness interventions during training through a reduction strategy. The second is a post-processing technique that adjusts output probabilities using plug-in estimation. Our theoretical analysis confirms that both approaches converge toward the optimal accuracy-fairness Pareto frontier. Empirical evaluations across several datasets validate the effectiveness of our proposed methods in successfully balancing predictive accuracy with fairness.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC