Trustworthy Recommender Systems for Multiple Stakeholders - Francesco Ricci, keynote at AVI' 26

Trustworthy Recommender Systems for Multiple Stakeholders Francesco Ricci, Free University of Bozen-Bolzano, Bolzano, Italy Keynote talk at AVI '26: International Conference on Advanced Visual Interfaces Venice Italy June 8 - 12, 2026 https://www.unive.it/avi2026/ Recommender systems (RSs) are personalized artificial intelligence (AI) tools routinely used by online platforms to promote algorithmically selected news, posts, music, travel, and videos. Today, generative AI techniques are impacting RS implementation and how users interact with these systems. With the introduction of generative AI, an RS’s core computational functionality can be implemented with a larger repertoire of techniques, and its capability to influence human choices and behaviour is further amplified. In essence, by exploiting user behavioural data and textual information (reviews and item descriptions), a machine learning (ML) algorithm is trained to identify and promote personalized items (products or services) for each user. The training objective is set by the platform owner and matches a desired operational goal, which normally relates to revenue growth and user engagement. In this keynote, after briefly discussing the values and risks of RSs, I will focus on the goal of making these systems more trustworthy, ensuring they better serve the various stakeholders interested in their functionality, especially the end users and the providers of the recommended items. The proposed solution leverages a more trustworthy evaluation method, which is also a mandatory requirement under European legislation (the Digital Services Act). A novel type of user/system interaction simulation framework will therefore be discussed. Compared to the standard train/test ML evaluation approach, this framework enables a better estimation of the potential longitudinal and multidimensional effects (positive and negative) that an RS can have on the decisions made by its users. The core idea is to more faithfully simulate users’ reactions to the recommendations produced by the tested RS by relying on a trustworthy algorithmic model of how users make their choices, i.e., select an item among the recommended ones. This approach departs from the standard ML testing mechanism that leverages testing data sampled from historical user/system interaction logs, which, unfortunately, were acquired when users interacted with unknown and different RSs. The application and usefulness of this evaluation approach will be exemplified in a particular case study: sustainable tourism management, specifically, how to mitigate overtourism and respect local communities. https://dl.acm.org/doi/10.1145/3811427.381...