Federated recommender system lets users control privacy and personalization
22 users, 8,807 titles: real-world test proves user-controlled AI recommendations work.
A new research paper from Manel Slokom and Alejandro Bellogin demonstrates a working federated recommender system that puts privacy control directly in users' hands. Unlike traditional recommenders that centralize data, this system keeps user interactions on-device and lets people actively choose between personalization or diversity-enhanced rankings. The live deployment ran for 53 days with 22 participants browsing a catalog of 8,807 titles, offering a rare real-world glimpse into how users behave when given explicit control over their recommendation objectives.
The results are compelling: users clicked more on personalized recommendations (65.37% CTR) than on diversity-enhanced ones (62.07%), showing a clear preference when given the choice. Participants rated their satisfaction with the control mechanisms at 3.93 out of 5 and adjusted settings 248 times during the study. The system also provided immediate feedback on how interactions affect recommendations, helping users develop an intuitive understanding of the trade-offs. This work challenges the assumption that privacy must come at the cost of personalization quality — and offers a practical blueprint for building recommender systems that respect both user autonomy and data privacy.
- Live 53-day deployment with 22 users and 8,807 titles proved federated recommender viability
- Users preferred personalization (65.37% CTR) over diversity (62.07%) when given explicit choice
- Participants made 248 settings changes and rated control satisfaction at 3.93/5
Why It Matters
Proves user-controlled, privacy-preserving recommendations are practical — a blueprint for AI systems that respect data sovereignty.