Efficient collaborative on-device recommender learning under resource and privacy constraints

Develop an efficient method for on-device recommender systems (DeviceRS) to incorporate collaborative information from other users while minimizing computation, storage, and data exposure, and accounting for heterogeneity (non-identically distributed data) across users' devices.

Background

The paper highlights that aggregating user–item interactions in centralized systems, while effective for leveraging collaborative information, is time-consuming, resource-intensive, and poses privacy and security risks. As a response, on-device recommender systems (DeviceRS) aim to train small models locally with lower computational and storage demands.

However, DeviceRS research is at an early stage and faces key unresolved issues: how to benefit from cross-user collaborative signals without violating privacy or incurring excessive communication and storage costs, and how to handle heterogeneous, non-identically distributed data residing on users' devices.