A new two-stage federated recommendation system that protects sensitive user data by keeping it on the mobile device, only sharing model updates during the training process.
This research introduces a novel two-stage federated recommendation system pipeline designed to serve personalized content on mobile devices while strictly adhering to privacy expectations. The core innovation involves separating non-sensitive user preference data from sensitive mobile context data, ensuring the sensitive information never leaves the device. Stage one runs a collaborative filtering model in the cloud using non-sensitive data to generate item shortlists. The second stage re-ranks these candidates locally on the device using sensitive mobile signals. This approach ensures that only model updates or gradients are transmitted, preserving user privacy.
The system was validated using MovieLens, UCI Human Activity Recognition, and a proprietary dataset, and a production-ready implementation is provided as a Kotlin Multiplatform library deployable on Android and iOS.