Files

Thesis_FVV_numeric_version.pdf
  • Restricted Access
  • Adobe PDF
  • 4.01 MB
Request a copy

Details

Authors
Supervisors
Fouss, François  
 ;
Saerens, Marco  
Abstract
(en) Artificial intelligence systems are increasingly embedded in organizational and everyday decision-making, raising ethical concerns that cannot be addressed through performance alone. Within the European framework for trustworthy AI, this thesis examines how risk mitigation methods can be developed and evaluated in supervised learning, with a focus on respect for human autonomy, fairness, and explicability, while treating the prevention of harm as a transversal concern. It combines methodological developments with offline and human-centered evaluations across several contexts. First, it investigates risks of filter bubbles in recommender systems through longitudinal simulations, showing that diversity and novelty depend on the interaction between algorithms and user behavior. Second, it develops fairness-aware classification methods to reduce demographic parity disparities, including an in-processing maximum-entropy logistic regression method and post-processing methods based on covariance constraints and a swapping mechanism. Experiments show that these methods can improve fairness while limiting losses in predictive performance. Third, the thesis studies reactions to fairness-oriented versus accuracy-oriented algorithms in human resource management decisions, showing that fairness-oriented algorithmic choices can increase moral approval when accuracy losses remain limited. Finally, it examines explicability in well-being prediction and shows that explanations, especially visual and interactive ones, improve user satisfaction. Overall, the thesis shows that risk mitigation in supervised learning requires both technical interventions and the evaluation of how these interventions are understood, accepted, and experienced by affected users.
Affiliations

Citations

  • APA
  • Chicago
  • FWB

Vancompernolle Vromman, F. (2026). Trustworthy AI : developing and evaluating risk mitigation methods in supervised learning. https://hdl.handle.net/2078.5/276704