FedPCL-CDR: A Federated Prototype-based Contrastive Learning Framework for Privacy-Preserving Cross-domain Recommendation

TL;DR

FedPCL-CDR tackles privacy-preserving cross-domain recommendation in settings with sparse overlapping users by introducing a federated, prototype-based contrastive learning framework. It leverages non-overlapping users through clustering to form representative prototypes and transfers knowledge via both local and global prototypes within an FL paradigm, with local differential privacy applied to prototypes to strengthen privacy. The method combines LightGCN-based graph representations, differential prototype learning, and dual-prototype contrastive losses to align embeddings across domains while protecting individual user data. Empirical results on Amazon and Douban tasks show state-of-the-art performance and favorable privacy-utility trade-offs, highlighting the practicality of prototype-based FL for PPCDR. The work also analyzes robustness to overlap ratio, ablations, and hyperparameters, providing insights for deploying privacy-aware cross-domain recommendations in real-world, data-scarce scenarios.

Abstract

Cross-domain recommendation (CDR) aims to improve recommendation accuracy in sparse domains by transferring knowledge from data-rich domains. However, existing CDR approaches often assume that user-item interaction data across domains is publicly available, neglecting user privacy concerns. Additionally, they experience performance degradation with sparse overlapping users due to their reliance on a large number of fully shared users for knowledge transfer. To address these challenges, we propose a Federated Prototype-based Contrastive Learning (CL) framework for Privacy Preserving CDR, called FedPCL-CDR. This approach utilizes non-overlapping user information and differential prototypes to improve model performance within a federated learning framework. FedPCL-CDR comprises two key modules: local domain (client) learning and global server aggregation. In the local domain, FedPCL-CDR first clusters all user data and utilizes local differential privacy (LDP) to learn differential prototypes, effectively utilizing non-overlapping user information and protecting user privacy. It then conducts knowledge transfer by employing both local and global prototypes returned from the server in a CL manner. Meanwhile, the global server aggregates differential prototypes sent from local domains to learn both local and global prototypes. Extensive experiments on four CDR tasks across Amazon and Douban datasets demonstrate that FedPCL-CDR surpasses SOTA baselines. We release our code at https://github.com/Lili1013/FedPCL CDR

Figures (3)

• Figure 1: The illustrations of knowledge transfer via overlapping and non-overlapping users.
• Figure 2: The framework of FedPCL-CDR. (1) Local Domain Learning: (a) Graph Representation Learning: Introducing LightGCN to learn comprehensive user and item embeddings. (b) Differential Prototype Learning: Clustering all user embeddings and applying LDP to obtain differential prototypes. (c) Prototype-based Contrastive Learning: Facilitating knowledge transfer across domains in a contrastive manner using both local and global prototypes. (2) Global Server Aggregation: It aggregates differential prototypes uploaded from multiple domains to generate local and global prototypes.
• Figure 3: Performance of different privacy budgets on Tasks 1 and 3.