Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation

Jing Liu, Lele Sun, Weizhi Nie, Peiguang Jing, Yuting SuAAAI 2024 · OpenReview

Abstract

Cross-Domain Recommendation (CDR) has been proven to effectively alleviate the data sparsity problem in Recommender System (RS). Recent CDR methods often disentangle user features into domain-invariant and domain-specific features for efficient cross-domain knowledge transfer. Despite showcasing robust performance, three crucial aspects remain unexplored for existing disentangled CDR approaches: i) The significance nuances of the interaction behaviors are ignored in generating disentangled features; ii) The user features are disentangled irrelevant to the individual items to be recommended; iii) The general knowledge transfer overlooks the user's personality when interacting with diverse items. To this end, we propose a Graph Disentangled Contrastive framework for CDR (GDCCDR) with personalized transfer by meta-networks. An adaptive parameter-free filter is proposed to gauge the significance of diverse interactions, thereby facilitating more refined disentangled representations. In sight of the success of Contrastive Learning (CL) in RS, we propose two CL-based constraints for item-aware disentanglement. Proximate CL ensures the coherence of domain-invariant features between domains, while eliminatory CL strives to disentangle features within each domains using mutual information between users and items. Finally, for domain-invariant features, we adopt meta-networks to achieve personalized transfer. Experimental results on four real-world datasets demonstrate the superiority of GDCCDR over state-of-the-art methods.

Introduction

Recommender systems frequently encounter the "data sparsity" challenge, where insufficient user interaction data in a specific target domain (e.g., Electronics) hinders accurate predictions. Cross-Domain Recommendation (CDR) addresses this by leveraging richer data from a source domain (e.g., Books). While disentangled representation learning—separating user interests into domain-invariant (universal) and domain-specific (unique) components—has shown promise, existing methods suffer from three primary limitations: interaction nuances are often ignored, disentanglement is typically item-agnostic, and knowledge transfer mechanisms are generally static and non-personalized.

Conceptual comparison of BCDR, existing DCDR, and the proposed GDCCDR framework.

As illustrated in Figure 1, traditional approaches often fail to account for the specific items involved in the disentanglement process, leading to "blurry" representations where universal and specific interests are not clearly separated. GDCCDR addresses these by introducing item-aware disentanglement and personalized transfer mechanisms.

Methodology: The GDCCDR Framework

The GDCCDR framework introduces four key technical innovations to refine cross-domain knowledge transfer:

  1. Adaptive Parameter-Free Filter (APFF): This component acts as a dynamic weight for graph convolutions, gauging the significance of each user-item interaction. It determines how much an interaction should contribute to domain-invariant versus domain-specific features based on the inherent structure of the interaction graph.
  2. Graph Disentanglement Module: Utilizing Graph Neural Networks (GNNs), the model propagates information through the user-item graph. The APFF ensures that features are disentangled from the initial stages of information propagation.
  3. Dual Contrastive Learning (CL) Constraints:
    • Proximate CL: Aligns the domain-invariant features of the same user across different domains, ensuring that universal preferences (e.g., a taste for high-quality goods) are consistent.
  4. Eliminatory CL: A novel constraint based on Mutual Information that ensures domain-specific features are highly relevant to their own domain while being effectively "eliminated" from the representation used for the other domain.
  5. Personalized Transfer via Meta-Networks: Rather than using a fixed transfer function, GDCCDR employs meta-networks to generate user-specific transfer matrices. This allows the system to adapt the transfer process based on individual user characteristics, such as their specific "transferability" between domains.
The complete GDCCDR architecture, including the Graph Disentanglement Module, Personalized Transfer, and Model Optimization components.

Experimental Evaluation

Datasets and Setup

The model was evaluated using four real-world cross-domain scenarios from the Amazon dataset: Sport & Phone, Sport & Cloth, Elec & Phone, and Elec & Cloth. These scenarios provide a rigorous testbed for handling varying levels of data sparsity and domain overlap.

Key Results

GDCCDR was compared against 11 baselines, including standard collaborative filtering (LightGCN) and state-of-the-art CDR models (BITGCF, DCCDR). The results highlight several key strengths:

  • State-of-the-Art Performance: GDCCDR consistently outperformed all baselines. Notably, on the Sport & Phone dataset, it achieved a nearly 8% improvement in Hit Ratio (HR@10) over the strongest baseline.
  • Robustness to Data Sparsity: The framework showed the most significant gains for users with very few interactions. This confirms that the personalized transfer and refined disentanglement effectively bridge the gap when local domain data is scarce.
  • Component Efficacy: Ablation studies confirmed that the removal of any core component—the meta-networks, the APFF, or the contrastive losses—resulted in a measurable performance decline, validating the integrated design of the framework.

Conclusion

GDCCDR advances cross-domain recommendation by moving beyond static, item-agnostic disentanglement. By incorporating adaptive filtering, dual contrastive constraints, and personalized meta-network transfer, the framework provides a more nuanced and effective way to leverage knowledge across domains. Future work will explore extending this framework to scenarios with non-overlapping users and optimizing the computational overhead associated with the meta-network training.

Summary updated

Abstract

Cross-Domain Recommendation (CDR) has been proven to effectively alleviate the data sparsity problem in Recommender System (RS). Recent CDR methods often disentangle user features into domain-invariant and domain-specific features for efficient cross-domain knowledge transfer. Despite showcasing robust performance, three crucial aspects remain unexplored for existing disentangled CDR approaches: i) The significance nuances of the interaction behaviors are ignored in generating disentangled features; ii) The user features are disentangled irrelevant to the individual items to be recommended; iii) The general knowledge transfer overlooks the user's personality when interacting with diverse items. To this end, we propose a Graph Disentangled Contrastive framework for CDR (GDCCDR) with personalized transfer by meta-networks. An adaptive parameter-free filter is proposed to gauge the significance of diverse interactions, thereby facilitating more refined disentangled representations. In sight of the success of Contrastive Learning (CL) in RS, we propose two CL-based constraints for item-aware disentanglement. Proximate CL ensures the coherence of domain-invariant features between domains, while eliminatory CL strives to disentangle features within each domains using mutual information between users and items. Finally, for domain-invariant features, we adopt meta-networks to achieve personalized transfer. Experimental results on four real-world datasets demonstrate the superiority of GDCCDR over state-of-the-art methods.
BibTeX
@inproceedings{liu2024graph,
  author = {Jing Liu and Lele Sun and Weizhi Nie and Peiguang Jing and Yuting Su},
  title = {{Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=GrVZu5xfxL},
}
Figure