LLM-enhanced multimodal fusion for cross-domain sequential recommendation

AbstractCross-Domain Sequential Recommendation (CDSR) predicts user behavior by leveraging historical interactions across multiple domains, capturing both intra- and inter-sequence item relationships. To further enhance the value of visual and textual data, we propose LLM-EMF, an innovative approach that incorporates Large Language Models(LLM) to enrich textual data and boosts recommendation performance by merging visual and textual information. Additionally, a multi-attention mechanism is designed to jointly learn single-domain and cross-domain preferences, effectively capturing complex user interests. Evaluations on four e-commerce datasets demonstrate that LLM-EMF outperforms existing methods in modeling cross-domain user preferences, highlighting the advantages of multimodal integration in sequential recommendation systems.