LAGOON: Achieving bounded individual fairness through classification frequency equalization

Abstract

One of the main concerns about using machine learning models for classification is algorithmic discrimination. Several works define different meanings of fairness to avoid or mitigate unfair classifications against minorities. The achievement of algorithmic fairness implies modifying training data, model operation, or outputs. Hence, the fair algorithm may modify the original classification. Generally, fairness means not discriminating against a person or a group. In a utopia, a system would classify every person or minority as privileged, which may decrease the utility of classification. We define λ-fairness, a relaxation of individual fairness designed to achieve fairness while maintaining utility with configurable parameters. We also propose a post-processing method that uses frequency equalization to achieve fairness in machine learning models by generalizing the outputs into frequencies. We used this flexible approach on LAGOON, an algorithm that achieves λ-fairness using frequency equalization. For experiments, we employ three benchmarks with different contexts to evaluate the quality of our approach. We compared our results to two baselines that aim to achieve fairness and minimize utility loss.