[Paper Review] JSCN: Joint Spectral Convolutional Network for Cross Domain Recommendation

Zhiwei Liu, Lei Zheng+3 more|arXiv (Cornell University)|Oct 18, 2019

Recommender Systems and Techniques38 references4 citations

TL;DR

This paper proposes JSCN, a Joint Spectral Convolutional Network that leverages multi-layer spectral convolutions on user-item graphs to capture high-order connectivity information and jointly learns domain-invariant user representations via a domain-adaptive mapping module. Experiments on 24 Amazon datasets show JSCN achieves 9.2% higher recall and 36.4% higher MAP than state-of-the-art methods, demonstrating superior cross-domain recommendation performance by effectively transferring knowledge across incompatible domains while preserving structural connectivity.

ABSTRACT

Cross-domain recommendation can alleviate the data sparsity problem in recommender systems. To transfer the knowledge from one domain to another, one can either utilize the neighborhood information or learn a direct mapping function. However, all existing methods ignore the high-order connectivity information in cross-domain recommendation area and suffer from the domain-incompatibility problem. In this paper, we propose a extbf{J}oint extbf{S}pectral extbf{C}onvolutional extbf{N}etwork (JSCN) for cross-domain recommendation. JSCN will simultaneously operate multi-layer spectral convolutions on different graphs, and jointly learn a domain-invariant user representation with a domain adaptive user mapping module. As a result, the high-order comprehensive connectivity information can be extracted by the spectral convolutions and the information can be transferred across domains with the domain-invariant user mapping. The domain adaptive user mapping module can help the incompatible domains to transfer the knowledge across each other. Extensive experiments on $24$ Amazon rating datasets show the effectiveness of JSCN in the cross-domain recommendation, with $9.2\%$ improvement on recall and $36.4\%$ improvement on MAP compared with state-of-the-art methods. Our code is available online ~\footnote{https://github.com/JimLiu96/JSCN}.

Motivation & Objective

• Address the data sparsity problem in recommender systems by transferring knowledge across domains.
• Overcome the limitations of existing methods that ignore high-order connectivity information in cross-domain recommendation.
• Solve the domain-incompatibility problem where user behaviors differ significantly across domains like movies and clothing.
• Jointly learn domain-invariant user representations using spectral convolutions and adaptive mapping to improve transferability.
• Demonstrate the effectiveness of multi-source domain learning and the impact of model architecture choices on performance.

Proposed method

• Apply multi-layer spectral convolutions on user-item bipartite graphs to extract high-order connectivity information using graph Fourier transforms.
• Use the graph Laplacian and its eigenvectors to define spectral representations that model non-linear interactions between users and items.
• Introduce a domain-adaptive user mapping module that projects domain-specific spectral vectors into a shared, domain-invariant user representation space.
• Formulate a joint optimization objective minimizing both in-domain reconstruction loss and cross-domain alignment loss to learn shared representations.
• Implement a linear or non-linear mapping function for the domain-adaptive module, with ablation showing linear mapping performs better due to reduced overfitting.
• Train the model end-to-end using stochastic gradient descent to optimize the joint loss on multiple source domains.

Experimental results

Research questions

• RQ1 Can spectral convolutions effectively capture high-order connectivity information in cross-domain recommendation?
• RQ2 Does learning a domain-invariant user representation improve recommendation performance across incompatible domains?
• RQ3 How does the inclusion of multiple source domains affect the performance of cross-domain recommendation models?
• RQ4 What is the impact of the domain-adaptive mapping module on knowledge transfer between domains with different behavioral patterns?
• RQ5 Does the choice of mapping function (linear vs. non-linear) significantly affect model generalization and performance?

Key findings

• JSCN achieves a 9.2% improvement in recall and a 36.4% improvement in MAP compared to state-of-the-art methods on 24 Amazon datasets.
• The domain-adaptive user mapping module significantly improves performance, with JSCN-β outperforming JSCN-α by up to 10% in MAP and 8% in recall.
• Using multiple source domains (e.g., Home and Kitchen + Health and Personal Care) improves performance by 37.2% on average compared to single-source learning.
• The combination of three source domains slightly underperforms two-source combinations, indicating that domain density and compatibility affect model performance.
• Linear mapping in the domain-adaptive module outperforms non-linear MLP mapping, likely due to reduced overfitting on low-dimensional spectral vectors.
• Spectral convolutions effectively model high-order connectivity, enabling better preference prediction—e.g., correctly ranking items based on multi-hop paths across domains.