Efficient Post-Processing for Equal Opportunity in Fair Multi-Class Classification

Ruicheng Xian, Han Zhao

Keywords: fairness, equal opportunity, multi-class, classification

TL;DR: We propose a fair post-processing method for equality of opportunity on multi-class problems.

Abstract: Fairness in machine learning is of growing concern as more instances of biased model behavior are documented while their adoption continues to rise. The majority of studies have focused on binary classification settings, despite the fact that many real-world problems are inherently multi-class. This paper considers fairness in multi-class classification under the notion of parity of true positive rates—an extension of binary class equalized odds—which ensures equal opportunity to qualified individuals regardless of their demographics. We focus on algorithm design and provide a post-processing method that derives fair classifiers from pre-trained score functions. The method is developed by analyzing the representation of the optimal fair classifier, and is efficient in both sample and time complexity, as it is implemented by linear programs on finite samples. We demonstrate its effectiveness at reducing disparity on benchmark datasets, particularly under large numbers of classes, where existing methods fall short.

Submission Number: 2943