Fairness-Aware Oversampling Algorithm Based on Distributions of Sensitive Attributes

Abstract

Given the growing utilization of machine learning algorithms in decision-making processes that impact individuals, it is crucial to ensure that predictions do not exhibit bias towards specific population subgroups, such as race or gender. Bias occurs when the likelihood of a positive outcome varies among different groups defined by these sensitive attributes. Research has demonstrated that this bias often stems from imbalanced datasets, where one class is significantly underrepresented compared to others. Understanding the nature of this imbalance, particularly the distribution of minority classes, is essential. This study introduces FAWOS, a Fairness-Aware oversampling technique designed to mitigate unfair treatment related to imbalances in sensitive attributes. FAWOS classifies data points based on their proximity to specific local neighborhoods defined by sensitive attributes, identifying those that are more challenging for classifiers to learn. To address dataset imbalance, FAWOS synthesizes new data points by oversampling from identified challenging data types. The impact of FAWOS on various classifiers is evaluated, assessing which ones are better equipped to handle imbalances related to sensitive attributes. Empirical findings indicate that FAWOS effectively enhances fairness in classifier outcomes without compromising overall performance. Key Words: Fairness-Aware Oversampling, Bias Mitigation, Imbalanced Datasets, Classifier Performance