Embedded Files
I analyze the long-term effects of current recommender system (RS) designs on the stakeholders involved (e.g., users, items, platforms, society) and identify interventions that can lead to better outcomes. Currently, I investigate emerging platforms and architectures that empower stakeholders to actively participate in the RSs' design and evaluation.
Poster_RecSys_SC.pdfInterpersonal Comparability
Interpersonal Comparability
Can we compare recommendation models based on the benefits they induce for stakeholders?
In this RecSys'26 work, we found the lack of interpersonal comparability to be an obstacle in doing so:
Many ranking-based metrics do not yield comparable benefits
Because of this, even Oracles appear unfair.
This "unfairness" doesn't capture a bias in the recommender system, but heterogeneity in the user population.
However, we found that using relative gains or Nash Social welfare can restore comparability. This is the poster we present at RecSys'26.
Page updated
Google Sites
Report abuse