Less data, more questions: fairness and accuracy under data minimization in recommender systems

Privacy laws now require data minimization, but its broader effects on recommender systems (RS) are still unclear. We systematically study how common minimization techniques reshape the three key RS goals—accuracy, user fairness, and provider fairness. Across multiple datasets and models we (i) measure performance shifts under data minimization strategies, (ii) pinpoint techniques that best balance the three objectives, and (iii) compare model robustness to data reduction. We find that while several strategies improve group-level consumer fairness, they often reduce accuracy and can even worsen provider fairness; the size of these trade-offs strongly depends on the chosen technique and model. Code and data are public at https://github.com/salvatore-bufi/ DataMinimizationFairness.