Speaker: Leonardo Tomazeli Duarte, University of Campinas (Unicamp, Brazil)
Title: Fairness-Aware Unsupervised Learning: Contributions to PCA and Data Aggregation
Location: SiMul meeting room, Faculté des Sciences et Technologies, Henri Poincaré Building, 4th floor
Abstract:
In this talk, we address fairness issues in unsupervised learning and data aggregation. First, we show that classical principal component analysis (PCA) may induce disparities in reconstruction errors across sensitive groups, leading to biased representations. A simple fairness-aware PCA approach is presented, based on a one-dimensional search that exploits the closed-form solution of PCA, significantly reducing group disparities with minimal loss in overall performance. Second, we discuss fairness in multicriteria decision analysis, introducing a framework for adjusting aggregation operators to produce fairer rankings. The approach relies on optimizing a fairness metric for both additive weighting and Choquet integral models, and its effectiveness is demonstrated on synthetic and real datasets.