Enhancing Dual-Target Cross-Domain Recommendation with Federated Privacy-Preserving Learning

Zhenghong Lin, Wei Huang, Hengyu Zhang, Jiayu Xu, Weiming Liu, Xinting Liao, Fan Wang

Abstract

Dual-target cross-domain recommendation (DTCDR) aims to alleviate data sparsity by leveraging knowledge across multiple domains. However, traditional DTCDR methods often require sharing sensitive user data, raising significant privacy concerns. In this paper, we introduce P2DTR, a novel federated learning framework designed for DTCDR that enhances recommendation accuracy while strictly preserving user privacy. P2DTR utilizes Private Set Intersection (PSI) to identify shared users and a prototype-based federated learning mechanism to exchange abstracted preference features rather than raw data or gradients. This approach is further bolstered by a Constrained Dominant Set (CDS) propagation mechanism that disseminates cross-domain knowledge to non-shared users. Extensive experiments on real-world datasets demonstrate that P2DTR consistently outperforms state-of-the-art baselines, particularly in cold-start scenarios.

Introduction

Modern recommendation systems frequently encounter the data sparsity problem, where a lack of user interactions hinders the algorithm's ability to learn accurate preferences. Dual-target cross-domain recommendation (DTCDR) addresses this by sharing information between two related platforms (e.g., Books and Movies). While effective, traditional DTCDR assumes platforms can directly access and match personal data of "shared users."

In the current landscape of strict privacy regulations like GDPR and CCPA, direct data sharing is often prohibited. Existing federated learning approaches attempt to solve this by sharing model gradients, but these remain vulnerable to "gradient inversion" attacks that can reconstruct raw data. P2DTR addresses these challenges by providing a framework that transfers valuable knowledge without compromising identifiable information.

Comparison of traditional DTCDR, Federated DTCDR, and the proposed P2DTR approach.

Methodology

The P2DTR framework operates through two primary stages: inter-client knowledge extraction and intra-client recommendation enhancement.

Inter-client Knowledge Extraction

To bridge domains privately, P2DTR first employs Private Set Intersection (PSI) based on RSA encryption. This allows two clients to identify shared users without revealing the identities of their exclusive users.

Instead of sharing raw embeddings or gradients, P2DTR utilizes Prototype-based Federated Learning. Clients learn "prototypes"—abstracted group features representing common preference clusters (e.g., "users interested in animation"). These prototypes are further protected using Differential Privacy (DP) by adding noise before they are sent to a central server for aggregation. This ensures that even if the server is compromised, individual user preferences cannot be reverse-engineered.

Intra-client Enhanced Recommendation

Once aggregated prototypes are received from the server, the Constrained Dominant Set (CDS) Propagation mechanism is used. This allows the "common knowledge" captured in the prototypes to be propagated through the local user-item interaction graph. Crucially, this benefits not just the shared users, but also the "local-only" users by refining their embeddings based on the broader cross-domain context.

The detailed architecture of the P2DTR framework, showing the flow from local graph convolution to prototype learning and server aggregation.

Experimental Evaluation

The researchers evaluated P2DTR on four real-world Amazon datasets across two federated scenarios: Game-Video and Cloth-Sport. Performance was measured using standard metrics including Mean Reciprocal Rank (MRR), Normalized Discounted Cumulative Gain (NDCG), and Hit Ratio (HR).

Key Results

  • Superior Accuracy: P2DTR consistently outperformed all state-of-the-art baselines. Notably, on the "Cloth" dataset, it achieved an 18.32% improvement in NDCG@10 over the best-performing existing model.
  • Cold-Start Performance: The framework showed significant advantages for users with few interactions, with MRR improvements ranging from 5.24% to 17.35% compared to other federated models.
  • Privacy-Utility Balance: Experiments confirmed that P2DTR maintains high recommendation utility even under the strict noise constraints required for robust Differential Privacy.

Conclusion

P2DTR represents a significant step forward in privacy-preserving recommendation. By replacing risky gradient sharing with abstracted prototype exchange and leveraging CDS propagation, it effectively bridges the gap between data utility and user privacy. Future work will explore scaling the framework to multi-client scenarios and incorporating heterogeneous data types such as text and images.

Summary updated

BibTeX
@inproceedings{lin2024enhancing,
  author = {Zhenghong Lin and Wei Huang and Hengyu Zhang and Jiayu Xu and Weiming Liu and Xinting Liao and Fan Wang},
  title = {{Enhancing Dual-Target Cross-Domain Recommendation with Federated Privacy-Preserving Learning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=ee5wkRm0ij},
}
Figure