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, Shiping Wang, Yanchao Tan

Abstract

Recently, dual-target cross-domain recommendation (DTCDR) has been proposed to alleviate the data sparsity problem by sharing the common knowledge across domains simultaneously. However, existing methods often assume that personal data containing abundant identifiable information can be directly accessed, which results in a controversial privacy leakage problem of DTCDR. To this end, we introduce the P2DTR framework, a novel approach in DTCDR while protecting private user information. Specifically, we first design a novel inter-client knowledge extraction mechanism, which exploits the private set intersection algorithm and prototype-based federated learning to enable collaboratively modeling among multiple users and a server. Furthermore, to improve the recommendation performance based on the extracted common knowledge across domains, we proposed an intra-client enhanced recommendation, consisting of a constrained dominant set (CDS) propagation mechanism and dual-recommendation module. Extensive experiments on real-world datasets validate that our proposed P2DTR framework achieves superior utility under a privacy-preserving guarantee on both domains.

Introduction

Cross-domain recommendation (CDR) is a critical technique for addressing the "data sparsity" problem, where recommendation systems lack sufficient interaction data to generate accurate suggestions. Dual-target CDR (DTCDR) extends this by simultaneously improving performance in two domains through the sharing of knowledge about overlapping users.

However, the practical deployment of DTCDR faces significant privacy hurdles. Regulations like GDPR and CCPA restrict the direct merging of databases from different domains. While standard Federated Learning (FL) offers a partial solution by sharing model gradients instead of raw data, it remains vulnerable to gradient inversion attacks. The core challenge lies in transferring cross-domain knowledge while maintaining a mathematically grounded privacy guarantee.

Comparison of DTCDR paradigms: (a) traditional DTCDR, (b) Federated DTCDR, and (c) the proposed P2DTR.

The P2DTR Framework

The proposed P2DTR framework addresses the privacy-utility trade-off through a two-stage process: inter-client knowledge extraction and intra-client recommendation enhancement.

Inter-Client Knowledge Extraction

To securely identify shared users between domains without exposing non-shared user identities, P2DTR utilizes an RSA-based Private Set Intersection (PSI) algorithm. Once the "bridge" users are identified, the framework employs Prototype-based Federated Learning. Instead of sharing raw data or sensitive model gradients, clients share "prototypes"—abstracted centroids representing user preference groups (e.g., users interested in "Comedy"). To further harden security, Differential Privacy (DP) is applied by adding controlled noise to these prototypes before they are aggregated by a central server. This ensures that no individual user's data can be reverse-engineered from the shared knowledge.

Intra-Client Enhanced Recommendation

After receiving the global aggregated prototypes, each domain must integrate this knowledge locally. This is achieved via Constrained Dominant Set (CDS) Propagation, which clusters local users around the global features. This mechanism acts as a bridge, injecting collaborative cross-domain relations into the local embedding space. Finally, a Dual-Recommendation Module uses graph-based collaborative filtering and Bayesian Personalized Ranking (BPR) loss to generate the final recommendations.

The overall architecture of the P2DTR framework.

Experimental Evaluation

The P2DTR framework was evaluated using large-scale Amazon datasets in two federated scenarios: Game-Video and Cloth-Sport. These datasets exhibit high sparsity (0.02% to 1.7% density), providing a rigorous test for cross-domain transfer capabilities.

Key Results

  • Performance Superiority: P2DTR consistently outperformed state-of-the-art baselines. Notably, on the "Cloth" dataset, it achieved an 18.32% improvement in NDCG@10 over the best baseline.
  • Cold-Start Mitigation: The model showed significant gains for users with few interactions, demonstrating the effectiveness of the cross-domain knowledge transfer.
  • Ablation Studies: Removing either the inter-client prototype sharing or the intra-client CDS propagation led to substantial performance drops, validating the necessity of both components.

Conclusion

P2DTR demonstrates that privacy-preserving constraints do not necessarily degrade recommendation utility. By sharing abstracted prototypes rather than gradients and utilizing CDS for local integration, the framework provides a robust, communication-efficient, and privacy-compliant solution for dual-target cross-domain recommendation. Future work will focus on scaling the framework to multi-domain scenarios and optimizing the computational overhead of the PSI and CDS components.

Summary updated

Abstract

Recently, dual-target cross-domain recommendation (DTCDR) has been proposed to alleviate the data sparsity problem by sharing the common knowledge across domains simultaneously. However, existing methods often assume that personal data containing abundant identifiable information can be directly accessed, which results in a controversial privacy leakage problem of DTCDR. To this end, we introduce the P2DTR framework, a novel approach in DTCDR while protecting private user information. Specifically, we first design a novel inter-client knowledge extraction mechanism, which exploits the private set intersection algorithm and prototype-based federated learning to enable collaboratively modeling among multiple users and a server. Furthermore, to improve the recommendation performance based on the extracted common knowledge across domains, we proposed an intra-client enhanced recommendation, consisting of a constrained dominant set (CDS) propagation mechanism and dual-recommendation module. Extensive experiments on real-world datasets validate that our proposed P2DTR framework achieves superior utility under a privacy-preserving guarantee on both domains.
BibTeX
@inproceedings{lin2024enhancing,
  author = {Zhenghong Lin and Wei Huang and Hengyu Zhang and Jiayu Xu and Weiming Liu and Xinting Liao and Fan Wang and Shiping Wang and Yanchao Tan},
  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