Comparative Study of Cross-Domain Recommendation Technologies: Focusing on JPEDET and SIEOUG Methods

Abstract

Against the backdrop of the digital information explosion, recommendation systems have become the core bridge connecting users with massive amounts of information. Cross-domain recommendation technology, which serves as a key means to tackle the issues of data sparsity and cold start faced in single domains, has gained significant attention during its development. This paper focuses on a core challenge in cross-domain recommendation—knowledge transfer in scenarios with no overlapping users. It systematically sorts out the mainstream technical routes in this field, including adversarial alignment, optimal transport, attribute alignment, and integrated innovative methods. Centering on the JPEDET method proposed at the AAAI Conference and the SIEOUG method proposed at The WebConf, the paper conducts an in-depth exploration of the differences between the two in terms of technical principles, applicable scenarios, and performance through theoretical analysis and experimental comparison. Experimental results on multiple sets of real cross-domain scenarios (such as book-music and movie-book) show that the JPEDET method, by integrating preference modeling of user ratings and reviews and a dynamic bidirectional embedding transportation mechanism, achieves the optimal performance in scenarios with no overlapping users. Compared with methods such as Temporal DomainAdaptive Recommendation(TDAR), Rectified Flow, and Collaborative Filtering with Attribution Alignment(CFAA), it achieves an average improvement in Root Mean Square Error(RMSE) and Mean Absolute Error(MAE) metrics. Additionally, this paper summarizes the challenges faced by current technologies and provides an outlook on future development directions, aiming to offer references for the research and application of cross-domain recommendation systems.