Publications

Machine learning models can exhibit error disparities across protected groups that fairness assessments may not reveal, e.g., due to lack of access to sensitive features or groundtruth labels, or failure to consider feature interactions. We investigate clustering as an approach to address such challenges. We experiment with a binary classifier trained on the OULAD dataset using sensitive and regular features, and with fairness assessment when limited information is available: no groundtruth labels or sensitive features. To further assess the impact of unavailable sensitive features, we also experiment with using sensitive features alone, together with regular features, and with increasing weights. We also systematically vary the clustering algorithm, distance metric, and data balancing strategy. Clustering quality is compared in terms of silhouette scores, error disparities among clusters, and whether clusters identify distinct social groups with specific sensitive features. We show that clustering results are greatly dependent on the choice of clustering algorithms, distance metrics, and data balancing strategy. Gower's distance outperformed Euclidean distance for mixed-type data, and K-Medoids was in most cases the more robust algorithm for fairness detection. These findings highlight the need for setting clear priorities, and for extensive feature engineering and parameter tuning, rather than using a one-size-fits-all approach.