Skip to content
Book Open access

FedTGC: Trajectory-Guided Graph Collaboration for Federated Recommendation

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 44 references

Abstract

Federated recommendation preserves user privacy by keeping interaction data on local clients, but this decentralization limits direct access to cross-user collaborative signals. Existing graph-based federated recommendation methods attempt to recover such signals by constructing client relation graphs. However, our empirical analysis shows that under a temporally faithful evaluation protocol, in which negative sampling and candidate construction are restricted to information available at the recommendation cutoff, client relations constructed by existing graph-based methods can become increasingly unreliable in practice. Consequently, clients may collaborate with mismatched neighbors, thereby introducing harmful collaborative signals and causing negative transfer. To address this problem, we propose FedTGC, a trajectory-guided graph collaboration framework for reliable federated recommendation. FedTGC represents each client by its local item embedding update trajectory, constructs a dynamic client relation graph from compact trajectory descriptors, and adaptively personalizes the calibration of collaboration strength across clients. This design enables effective knowledge sharing among similar clients. Extensive experiments on five real-world datasets show that FedTGC consistently outperforms representative federated recommendation baselines under the Honest Setting. Further analyses reveal that the commonly adopted Original Setting can substantially overestimate recommendation performance, highlighting the importance of temporally faithful evaluation and validating the effectiveness of trajectory-guided collaboration. The source code is available at https://github.com/Charleston77/FedTGC.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.