arXiv paper Beyond Centralization reports a 53-day federated recommender deployment with 22 users and 8807 titles. Users achieved 65.37% CTR on personalization vs 62.07% on diversity-enhanced ranking when given explicit control.
Key facts
- 53-day deployment with 22 participants
- 8807 titles in the recommendation catalog
- 65.37% CTR for personalization vs 62.07% for diversity
- 3.93/5 user satisfaction with control mechanisms
- 248 settings changes recorded during the study
The paper, submitted to arXiv on April 10, 2026, presents a live federated recommender system that keeps user data on-device while allowing users to switch between personalization and diversity-enhanced ranking objectives. Over 53 days, 22 participants made 248 settings changes and rated the control mechanisms 3.93/5 satisfaction. The system maintained competitive CTR against typical centralized approaches, with personalization winning out when users explicitly chose it.
Why this matters more than the paper suggests
The result challenges a core assumption in the recommender systems community: that personalization quality inevitably degrades under federated constraints. Here, user-controlled federated recommendations not only matched but slightly exceeded typical centralized CTR baselines (65.37% vs 62.07%). The key enabler was giving users real-time feedback on how their choices affected recommendations, which drove engagement and learning. This suggests that the privacy-utility tradeoff may be overstated when users are active participants rather than passive data sources.
How the system works
The architecture uses a standard federated averaging approach with a twist: each device maintains a local model that can be tuned toward personalization or diversity via a user-adjustable slider. The server aggregates only encrypted gradient updates, never raw user data. The catalog of 8807 titles spans multiple genres, and the system logs every interaction for post-hoc analysis.
Limitations and open questions
The study's small sample (22 participants) limits statistical power. The paper does not disclose participant demographics or recruitment methods, making generalizability uncertain. Additionally, the 53-day window may not capture long-term drift in user preferences or system performance. The authors acknowledge these limitations and call for larger-scale deployments.
What to watch
Watch for follow-up studies with larger cohorts (100+ users) and longer durations (6+ months) that would test whether the CTR gains hold under real-world scale. Also watch for integration with existing federated learning frameworks like TensorFlow Federated or PySyft, which would lower the barrier to replication.
What to watch
Watch for larger-scale replications (100+ users, 6+ months) that would validate the CTR gains. Also watch for integration with TensorFlow Federated or PySyft, which would enable broader adoption of user-controlled federated recommendations.
