Meta-Learning Based Interest Selection for Cross-Domain Recommendation
Year: 2026
Type: conference-paper
Abstract: The user cold-start problem refers to the performance degradation of recommendation systems caused by data sparsity. By transferring information across related domains, cross-domain recommendation aims to improve recommendation effectiveness in the target domain. However, most related works transfer knowledge from the perspective of global shared mappings or user-level representation transfer, overlooking the differences among various interest dimensions during crossdomain migration. Furthermore, existing methods tend to treat all interests equally, which weakens the model's utilization of critical cross-domain information. To address these issues, we propose a Meta-learning based Interest Selection for Cross-Domain Recommendation (MISCDR). The model generates multi-interest vectors in each domain through a multi-interest extraction module, and employs a cross-domain interest selection module to perform relevance weighting on source domain interests. Subsequently, meta-networks are leveraged to construct multi-interest bridge functions, generating user representations under different interest dimensions across domains. Finally, the information from the source and target domains are fused through a dynamic fusion gating network. Experimental results on multiple cross-domain tasks demonstrate that the proposed MISCDR method achieves significant performance improvements over various baseline methods.
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204145453