.png){kind=icon}
.png){kind=icon}
Return
Guided diffusion cross-domain recommendation with consistency alignment
.png){kind=icon}
PRE
AI
B
X
DOI：10.1007/s00521-026-12427-y.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Cross-domain recommendation (CDR) represents a key area of study within the realm of recommendation systems, focusing on addressing the challenge posed by the scarcity of user interaction data in the target domain. Although current CDR approaches have achieved some advancements in alleviating the cold-start problem for users, they continue to face difficulties such as inadequate knowledge transfer and limited generalization capabilities of models. This paper presents a $$\underline{\hbox {G}}$$ uided $$\underline{\hbox {D}}$$ iffusion $$\underline{\hbox {C}}$$ ross- $$\underline{\hbox {D}}$$ omain $$\underline{\hbox {R}}$$ ecommendation framework based on $$\underline{\hbox {C}}$$ onsistency $$\underline{\hbox {a}}$$ lignment (GDCDR-CA). This framework implements cross-domain knowledge transfer using the Target Information-Guided Diffusion Module (TIGDIM) while enhancing the efficacy of this transfer through the Consistency Alignment Module (ALM) and the Restart Random Walk Module (RWRM). More specifically, TIGDIM utilizes a conditional reverse diffusion process to create feature representations for users in the target domain. ALM guarantees that the embedding vectors produced align with the traits of target domain users, while RWRM enhances the accuracy of knowledge transfer by incorporating information from neighboring entities. Experimental results derived from the Amazon review dataset indicate that GDCDR-CA surpasses existing methods significantly across different cold-start levels in cold-start situations. Additionally, ablation analyses validate the contributions of each component, reinforcing the advantages of GDCDR-CA in improving cross-domain recommendation outcomes.
Keywords:
User cold start
Guided diffusion model
Consistency alignment
Cross-domain recommendation
Journal
IF:
4.5
Papers:
729
Citations:
3.2W
Organization
Cited Papers