Stavroula Eleftheraki

Abstract

Recommender systems increasingly influence the information, services, and opportunities people encounter online, making fairness a critical requirement. While substantial research has focused on designing fairness-aware recommenders, fairness is often evaluated under the assumption that the underlying interaction data are fixed, overlooking how changes to these data can influence recommendation fairness. This thesis studies fairness in recommender systems from two complementary perspectives: fairness optimization through recommendation algorithm design and fairness manipulation through interaction-data interventions.

We first address fairness at the algorithmic level by developing efficient neighborhood-learning methods for collaborative filtering that improve consumer and provider fairness while offering different trade-offs between fairness and recommendation accuracy.

We then investigate the fragility of recommendation fairness under adversarial data interventions. First, we study fairness attacks based on fake-user profile injection, demonstrating that both conventional and fairness-aware recommenders can experience substantial fairness degradation, with fairness-aware models sometimes amplifying the effect. Second, we consider a different threat model in which an adversary manipulates the interactions of real users. We formulate this setting as a budget-constrained optimization problem that strategically adds or removes user–item interactions to alter the exposure of opportunities across demographic groups.

Overall, this thesis demonstrates that fairness should be viewed not only as an objective to optimize, but also as a property whose robustness and fragility must be understood. These findings highlight the importance of designing recommender systems whose fairness remains robust under adversarial manipulation of interaction data.