A comparative study of pre-processing algorithms for fair classification in few labeled data context

In previous research, the predominant approach for addressing fairness concerns in machine learning was rooted in supervised methods. However, this approach faced real-world challenges, such as obtaining trustworthy ground truth data and the potential for human bias within that data. To overcome these obstacles, we turned to the realm of semi-supervised learning (SSL) that uses unlabeled data to achieve a better trade-off between accuracy and discrimination. We aim to use SSL to generate pseudo-labels for the remaining unlabeled data, thereby enhancing supervised model training. In our study, we harnessed the power of two widely recognized semi-supervised techniques as pseudo-labelers with pre-processing mitigation techniques to construct fair classification models. So we conducted a systematic examination of various pre-processing techniques which included suppression of protected attribute, different ways of balancing, and representation learning algorithms like Correlation Remover and Disparate Impact Remover. Results in five datasets with diverse protected attributes, such as sex, race, and age, show improvement in fairness metrics like Disparate Impact and Average Absolute Odds Difference, and in some cases even in classification metrics like Area under the ROC Curve. This approach underscores the potential of combining semi-supervised learning with fairness-focused pre-processing techniques to create more equitable and effective machine learning models.