A Neighbor-based Expansion Approach to Graph Federated Recommendation

Federated recommendation offers a promising approach to addressing privacy and data-sharing challenges in recommendation systems. When combined with Graph Neural Networks (GNNs), it can effectively capture complex high-order interactions between users and items. However, in cross-device federated settings, where each user corresponds to a single device, the limited high-order depth of local interactions reduces GNN effectiveness in capturing global graph structure. User expansion techniques partly alleviate this limitation but often rely on third-party servers, introducing data duplication, privacy risks, and inefficiencies. To address these limitations, we propose a drop-in replacement for the expansion step of graph federated recommenders, a user expansion based on secure nearest-neighbour search over user embeddings, requiring neither shared interactions nor a trusted third-party server. Substituting only the expansion step, and leaving the rest of the pipeline untouched, we match or exceed shared-interaction expansion on four benchmark datasets while adding a fixed budget of $k$ users per client instead of an uncontrolled fraction of the dataset, which bounds the size of every local graph and therefore the recurring cost of local GNN training.