[Paper Review] Graph Factorization Machines for Cross-Domain Recommendation

Dongbo Xi, Fuzhen Zhuang+4 more|arXiv (Cornell University)|Jul 12, 2020

Recommender Systems and Techniques35 references5 citations

TL;DR

This paper proposes Graph Factorization Machines (GFM), a novel GNN model that leverages Factorization Machines to effectively aggregate multi-order neighborhood interactions in graph-structured data for improved recommendation performance. It further introduces a general cross-domain framework applicable to GFM and other GNNs, significantly enhancing performance on sparse data across multiple real-world datasets.

ABSTRACT

Recently, graph neural networks (GNNs) have been successfully applied to recommender systems. In recommender systems, the user's feedback behavior on an item is usually the result of multiple factors acting at the same time. However, a long-standing challenge is how to effectively aggregate multi-order interactions in GNN. In this paper, we propose a Graph Factorization Machine (GFM) which utilizes the popular Factorization Machine to aggregate multi-order interactions from neighborhood for recommendation. Meanwhile, cross-domain recommendation has emerged as a viable method to solve the data sparsity problem in recommender systems. However, most existing cross-domain recommendation methods might fail when confronting the graph-structured data. In order to tackle the problem, we propose a general cross-domain recommendation framework which can be applied not only to the proposed GFM, but also to other GNN models. We conduct experiments on four pairs of datasets to demonstrate the superior performance of the GFM. Besides, based on general cross-domain recommendation experiments, we also demonstrate that our cross-domain framework could not only contribute to the cross-domain recommendation task with the GFM, but also be universal and expandable for various existing GNN models.

Motivation & Objective

• Address the limitation of existing GNNs in capturing complex multi-order feature interactions due to simplistic aggregation mechanisms like mean or pooling.
• Overcome data sparsity in recommender systems by enabling effective cross-domain knowledge transfer using graph-structured data.
• Develop a generalizable cross-domain recommendation framework that is compatible with various GNN models, not limited to a specific architecture.
• Enhance recommendation performance by jointly learning domain-shared and domain-specific representations using shared node initialization and shared graph parameters.
• Demonstrate the effectiveness of the proposed GFM and cross-domain framework across diverse real-world datasets and multiple GNN baselines.

Proposed method

• Propose Graph Factorization Machine (GFM), which uses Factorization Machines to aggregate second-order neighbor messages, enabling effective modeling of feature interactions in sparse, high-dimensional user-item interactions.
• Stack multiple GFM layers to capture higher-order neighborhood information through iterative message passing, improving representation learning beyond first-order interactions.
• Design a general cross-domain recommendation framework that initializes source and target domain nodes using shared representations (e.g., shared users or items), enabling transfer of domain-shared knowledge.
• Train the model by jointly optimizing domain-specific graph representations using the graph topology of each domain, while sharing graph parameters across domains to encourage consistency.
• Combine domain-shared and domain-specific features at inference time to produce final predictions, allowing the model to benefit from both transferable and unique patterns.
• Apply the framework to various GNN models (GCN, GAT, GraphSAGE) to demonstrate its universality and expandability beyond GFM.

Experimental results

Research questions

• RQ1 Can a GNN model that leverages Factorization Machines better capture multi-order feature interactions in graph-structured recommender systems compared to standard GNNs?
• RQ2 How effective is the proposed cross-domain framework in improving recommendation performance when transferring knowledge between domains with sparse, graph-structured data?
• RQ3 To what extent does the proposed framework generalize across different GNN architectures such as GCN, GAT, and GraphSAGE?
• RQ4 Does shared node initialization significantly improve performance in cross-domain recommendation tasks compared to using only domain-specific representations?
• RQ5 Can the combination of domain-shared and domain-specific features lead to better performance than using either alone in cross-domain settings?

Key findings

• The proposed GFM model achieves state-of-the-art performance on four real-world cross-domain recommendation datasets, outperforming multiple strong baselines, including FM, GCN, and GAT.
• CD-GFM (cross-domain GFM) improves HR@1 by 25.2% on average and HR@10 by 10.5% on average compared to the base GFM model, demonstrating the effectiveness of the cross-domain framework.
• The ablation study confirms that the shared node initialization component contributes significantly to performance, with p-values < 0.05 in all tested datasets, indicating statistical significance.
• The general cross-domain framework improves performance across 40 different tasks (4 dataset pairs × 10 models), showing consistent gains across GCN, GAT, and GraphSAGE variants.
• The improvement is most pronounced for GCN, likely due to its weaker performance in the single-domain setting, highlighting the framework’s ability to correct for weak base models.
• The framework is robust and effective even when using simple aggregators like mean and pooling in GraphSAGE, suggesting broad compatibility with diverse GNN architectures.