(New) 11 September 2026, Computational Social Choice Seminar, Stefania Ionescu

Abstract

Recommender systems mediate access to information, jobs, housing, and services, shaping outcomes for users, content providers, platforms, and society at large. To do so, these systems train models that predict item relevance for each user. Traditional frameworks for deciding which model is best rely on a single metric to assess user satisfaction and choose the model that maximizes the average across users. From a social choice perspective, this corresponds to a utilitarian welfare aggregation rule applied only to the user side. Yet such rules are only meaningful when individual utilities are comparable across agents. But is it fair to assume they are? In this talk, I present both analytical and empirical evidence that several user satisfaction metrics used in practice do not necessarily yield comparable utility measures across individuals. Ignoring this evidence and using utilitarian welfare to aggregate utilities can lead to a different recommendation model being selected and to misleading fairness evaluations, in which even an Oracle may appear unfair. We find that considering relative gains can restore comparability for fairness analysis, while Nash social welfare provides a natural aggregation rule when social evaluation must remain invariant to stakeholder-specific rescaling. I will end the talk discussing open questions at the intersection of social choice and recommender systems.

For more information on the Computational Social Choice Seminar, please consult https://staff.science.uva.nl/u.endriss/seminar/.