Federated Graph Learning for Cross-Domain Recommendation

Summary

This paper introduces an innovative federated CDR framework with two key modules tailored for privacy preserving and negative transfer. For privacy, it presents a solid theoretical guarantee. For negative transfer, it generates domain attentions by virtual social links and conduct a fine-tuning stage to filter source domain knowledge. With the comprehensive empirical evaluation, it shows the superiority of the proposed framework.

Strengths

1. Innovation: Utilizing HVH structure and Gaussian noise to ensure all-round privacy, GAT with virtual links to generate domain attentions and comprehensive objective function with prediction loss, mapping loss and social regularization loss seems technically sound. 2. Theoretical analysis: The paper provides a robust theoretical analysis. The detailed explanation concerning protecting transferred high-order embeddings from inference attacks strengthen significantly strengthen the paper's theoretical framework. 3. Generally, the proposed method is complex but this paper is well-organized and easy to follow. 4. Experimental demonstrations: The authors have conducted extensive experiments to validate the effectiveness of the proposed method, including the performance comparison, the ablation study, the privacy budget study and an additional dual-domain study. The compare of overall performance validates the superiority of proposed model. The ablation study validates the effectiveness of two key modules and the privacy budget study demonstrate the balance of privacy preservation and model performance. The dual-domain study proves that the proposed model can also cope with traditional scenarios.

Weaknesses

1. In Figure 2, the use of the terms "extended" and "expansion" appears to be inconsistent. Consistent use of terminology throughout the paper will aid in the reader's comprehension and the overall clarity of the presentation 2. The authors have conducted extensive experiments on the Amazon dataset as detailed in the experimental section. However, as the authors themselves acknowledge in the limitation section, the generalization performance of their framework has not been sufficiently validated. Given this, and despite the fact that the experiments may not align with the authors' definition of BD-CDR, it is recommended that the authors consider conducting experiments on other datasets (e.g., Douban) to further verify the generalizability of their framework. This additional testing would strengthen the paper's claims regarding the robustness of the proposed approach.

Questions

1. From the perspective of privacy, DP is widely used in previous method. What are the privacy innovations in this paper? What is the unique technical contribution between this work and existing works on privacy-preserving CDR?

Rating

7

Confidence

5

Soundness

4

Presentation

4

Contribution

4

Limitations

Please refer to weaknesses.

Summary

This paper presents a novel federated framework of CDR, FedGCDR, for privacy preserving and negative transfer between domains. Its key strengths include a solid theoretical foundation analyzing the DP-based privacy preserving and a novel and dynamic attention generation method to mitigate negative transfer. By tackling both privacy concern and potential negative transfer problem, this work makes a valuable contribution to the field.

Strengths

1. Tackling two Critical Issues for cross-domain recommendation: The paper addresses crucial challenges for privacy preservation and potential negative transfer. By following a horizontal-vertical-horizontal FL pipeline and adopting the Gaussian mechanism, the paper ensures the privacy of both intra-domain and inter-domain. By the graph expansion, the paper generates domain attention to filter harmful info from multiple source domains. Addressing these two issues is essential for fostering a wide range of participation and maintaining a healthy balance between domain involvement and model performance, ultimately contributing to the overall sustainability and scalability of CDR. 2. Novel Approach: The proposed FedGCDR approach is novel in several aspects. First, the problem formulation itself, which focuses on mitigating negative transfer under the privacy constraint, is a departure from existing methods that directly transfer with the consumption of data sparsity and well-chosen domains. Second, the theoretical analysis of the reliability of the Gaussian mechanism protecting the high-order embeddings provides valuable insights and paves the way for a secure transfer between domains. Third, the incorporation of GAT as a tool for both mining domain knowledge and generating dynamic domain attention to filter potentially harmful information is a novel method to address the negative transfer issue, which has been a long-standing challenge in CDR. 3. Theoretical Guarantees and Practical Considerations: The paper presents a solid theoretical foundation by analyzing the Gaussian mechanism to protect high-order embeddings output by GAT. Furthermore, by combining horizontal fl with vertical fl, knowledge transfer down to edge devices is realistic and avoids the additional overhead required for the vertical process. 4. Comprehensive Empirical Evaluation: The paper's strength lies in its comprehensive empirical evaluation across sixteen widely used domains of the Amazon dataset, spanning various sub-datasets of different domain numbers. Besides the overall model performance comparison, the paper compares the ability of mitigating negative transfer with defined concept ‘soft negative transfer’ and ‘hard negative transfer’. The target domain setting with different data quality further demonstrates the generalization of the model and the irrationality of the direct transmission of the previous method.

Weaknesses

1. The paper addresses the broader-domain CDR involving more than three domains. While multi-domain CDR is undoubtedly significant and has been the subject of extensive research, the emphasis on the number "three" is noted. Could the authors clarify the significance of this number in their work? Does this particular emphasis distinguish their research from existing studies? If not central to the research, it is suggested that the authors reconsider the emphasis on this number to avoid potential misunderstandings or unnecessary focus. 2. By introducing multiple social regulation terms, yet the authors have chosen one specific term (as shown in Equation 9). It would be beneficial for the authors to provide an explanation for this selection. Is this choice based on theoretical analysis, empirical results, or comparative study with existing literature? Clearly articulating the basis for this choice will help readers understand its importance and applicability.

Questions

1. This article is mainly from the user's perspective to solve the privacy problem and whether the proposed framework can be extended to more realistic company-level applications. 2. Negative transfer is an important problem in CDR as well as transfer learning. However, facing the ‘when to transfer’ problem [1] in CDR, it seems to be a potential method to stop the negative transfer when it occurs. Why cannot we adopt this simple method to avoid such complex attention generation progress? [1] Pan S J, Yang Q. A survey on transfer learning[J]. IEEE Transactions on knowledge and data engineering, 2009, 22(10): 1345-1359.

Rating

8

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Please refer to the weak points.

Summary

The paper proposes a novel federated graph learning framework, FedGCDR, aimed at addressing the challenges of privacy and negative transfer (NT) in Broader-Source Cross-Domain Recommendation (BS-CDR) scenarios. The framework includes two key modules: the positive knowledge transfer module and the positive knowledge activation module. These modules ensure privacy preservation and mitigate NT by employing differential privacy and feature mapping techniques, followed by graph expansion and fine-tuning in the target domain. The framework is validated through extensive experiments on the Amazon dataset, demonstrating its superior performance over existing methods.

Strengths

1. The paper considers privacy preservation and negative transfer (NT) challenges under a more generic scenario of Broader-Source Cross-Domain Recommendation (BS-CDR). 2. The framework's design is modular, allowing for easy adaptation and potential integration with other recommendation system components. 3. The experiments cover 16 popular domains from the Amazon dataset and demonstrate that FedGCDR outperforms state-of-the-art methods in terms of recommendation accuracy.

Weaknesses

1. Despite the overall clarity of the writing, there are several clerical mistakes that could easily mislead the reader. 2. While the results on the Amazon dataset are promising, it is unclear how well the model generalizes to other datasets or real-world scenarios with different characteristics. 3. The authors emphasize that the contribution lies in provides high-quality BS-CDR recommendations while safeguarding both user privacy and domain confidentiality. However, the experimental section is completely devoid of any discussion on privacy preservation.

Questions

1. About the clerical error in writing, such as in line 105, should "privacy of individual users (inter-domain privacy)" be "intra-domain privacy"? The line 166 “we learn learning…”, and the line 584 “FedCT’s HR@10 performance is better on Amazon-4@CDs than on Amazon-8@CDs”. But according to Figure 6, it should be the Amazon-8@CDs that outperform the Amazon-4@CDs for FedCT's HR@10. 2. In Section 3.1, the authors assumed that users are partially overlapping between domains, so how do non-overlapping users perform graph expansion in the target domain? 3. The improvement in recommendation performance of the proposed framework is demonstrated in the experimental section, but why another important contribution on privacy preservation is not discussed? 4. Does FedGCDR-DP in Tables 1 and 2 refer to another variant method? What is the difference with the proposed FedGCDR? This is not explained relevantly in the paper. 5. Are the Amazon-4, Amazon-8 shown in Table 4 randomly divided? Or is there another domain selection strategy?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes. The authors have addressed the limitations.

Summary

To solve the privacy issue and negative transfer phenomenon in the cross-domain recommendation, the authors propose a novel framework named FedGCDR. Following the HVH pipeline, two key modules collaboratively transfer positive knowledge and filter the negative interference from source domains. In the experimental part, real world data in sixteen domains validated the model's performance.

Strengths

1. Novelty. Dynamic attention mechanism in the end-to-end framework ensures an automated domain filtering mechanism. The HVH pipeline makes the training of multi-domain scenes efficient, coupled with differential privacy technology, ensuring the privacy of the training process. 2. Theory. This paper adopts the Gauss mechanism guarantee on the middle embedding to defend against inference attacks and proves its validity by theory. 3. Experiments. Based on the Amazon data set, the author carried out a detailed experiment and analyzed the coping ability of various methods to the phenomenon of negative transfer from two perspectives of soft negative transfer and hard negative transfer, thus verifying the superiority of the proposed method. Based on verifying the effectiveness of the two modules, the detailed ablation experiment fully reveals the influence of the two modules on domains of different data quality, which is very convincing.

Weaknesses

1. The introduction and related works sections of the paper introduce two pivotal concepts: intra-domain privacy and inter-domain privacy. However, the definitions provided for these concepts appear to be inconsistent, which may obscure the paper's motivation regarding privacy concerns. For the paper to effectively convey its contributions and significance, it is crucial that these two concepts are clearly and uniformly defined. 2. The authors have chosen to emphasize the "Broader-Domain CDR" scenario, which involves more than three domains when defining the problem. While this distinction is noted, it is not immediately clear why the term "multi-domain CDR" is not used to introduce the framework, as it seems to be a more commonly accepted term in the literature.

Questions

In practical application, heterogeneity exists between different domains, and it may not be appropriate to use the same embedding dimension. Is there a possible solution?

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Please refer to the weak points.