Federated Graph Learning for Cross-Domain Recommendation

TL;DR

FedGCDR is proposed, a novel federated graph learning framework that securely and effectively leverages positive knowledge from multiple source domains that significantly outperforms state-of-the-art methods on popular domains of the Amazon dataset.

Abstract

Cross-domain recommendation (CDR) offers a promising solution to the data sparsity problem by enabling knowledge transfer across source and target domains. However, many recent CDR models overlook crucial issues such as privacy as well as the risk of negative transfer (which negatively impact model performance), especially in multi-domain settings. To address these challenges, we propose FedGCDR, a novel federated graph learning framework that securely and effectively leverages positive knowledge from multiple source domains. First, we design a positive knowledge transfer module that ensures privacy during inter-domain knowledge transmission. This module employs differential privacy-based knowledge extraction combined with a feature mapping mechanism, transforming source domain embeddings from federated graph attention networks into reliable domain knowledge. Second, we design a knowledge activation module to filter out potential harmful or conflicting knowledge from source domains, addressing the issues of negative transfer. This module enhances target domain training by expanding the graph of the target domain to generate reliable domain attentions and fine-tunes the target model for improved negative knowledge filtering and more accurate predictions. We conduct extensive experiments on 16 popular domains of the Amazon dataset, demonstrating that FedGCDR significantly outperforms state-of-the-art methods.

Figures (7)

• Figure 1: (a) In order to obtain accurate recommendations in the Books domain, we aim to exploit user preferences (i.e., knowledge of external domains should be fully utilized, e.g. Movie, Toys, and Games domains). However, with the influence of lossy privacy-preserving techniques, the results of the transfer could be negative (e.g., the Music domain with low-quality data). (b) There is a diminishing marginal effect on the growth rate of the model performance with pure positive knowledge, while NT accumulates with an increasing number of source domains. Consequently, the performance of existing methods declines and is worse than that of a single domain model.
• Figure 2: An overview of FedGCDR. It consists of two key modules and follows a HVH pipeline: (1) Source Domain Training (Horizontal FL): ① Each source domain maintains its graph attention network (GAT)-based federated model. (2) Positive Knowledge Transfer Module (Vertical FL): ② Source domain embeddings are extracted from GAT layers and perturbed with Gaussian noise. ③ The multilayer perceptron aligns the feature space of source domain embeddings and target domain embeddings. (3) Positive Knowledge Activation Module (Horizontal FL): ④ Local graph is expanded with source domain embeddings. ⑤ Enhanced federated training of the target domain is achieved through the expanded graph. ⑥ The target domain maintains its GAT-based federated model. ⑦ The target domain freezes the GAT layer and fine tunes the model.
• Figure 3: Illustration of target domain graph expansion. The virtual users are constructed with the source domain embeddings from the Movie domain and the Music domain. The attentions generated by social links to the virtual user can be regarded as the domain attentions.
• Figure 4: Illustrations of negative transfer on HR@5 and NDCG@5. Metric values lower than single-domain (dotted line and red area) mean severe negative soft negative transfer. The figure on HR@10 and NDCG@10 is shown in Appendix D.1.
• Figure 5: Ablation study on Amazon-16@CDs and Amazon-16@Books.