Generalization of Knowledge Transfer with User Reviews for Cross-Domain Recommendation

Cross-domain recommendation systems have demonstrated the potential to address data sparsity and cold-start problems. However, current approaches primarily rely on domain-shareable attributes, such as overlapping user bases or identical contexts, to facilitate knowledge transfer, limiting their generalizability without these elements. To overcome these limitations, we propose exploiting review texts, which are ubiquitous across most e-commerce platforms. Our model, termed SER, incorporates three distinct text analysis modules, guided by a single discriminator to achieve disentangled representation learning. We introduce a novel optimization strategy that not only improves domain disentanglement but also minimizes the transfer of adverse information from the source domain. Furthermore, we have expanded our model’s encoding network from a single to multiple domains, enhancing its efficacy for review-based recommendation systems. Through comprehensive experiments and ablation studies, we establish that our approach is more efficient, robust, and scalable than existing single and cross-domain recommendation methods. This paper is an extension of our prior work from CIKM ’22 [1], offering additional insights, further experimental results, and a comprehensive theoretical analysis.