LeCDSR: Large language model enhanced cross-domain sequential recommendation

As large language models (LLMs) have shown great performance in natural language processing, research on applying them to recommendation systems has emerged. LLMs’ strong understanding, reasoning, and extensive world knowledge can supplement the missing semantic information in recommendation systems. Existing LLM-enhanced recommendation systems face challenges in extracting and leveraging features, lack of sufficient utilization of LLMs’ capabilities to capture user interests. In this paper, a novel algorithm, Large Language Model enhanced Cross-Domain Sequential Recommendation, LeCDSR is proposed. LeCDSR generates cross-domain user profile embeddings through LLMs to transfer user preference information across domains. It also uses a semantic fusion layer to integrate semantic and ID embeddings, addressing the limitations of traditional sequential recommendation models. Furthermore, LeCDSR employs a contrastive loss function to better align the feature spaces of LLMs and recommendation models, improving recommendation performance in cross-domain scenarios. LeCDSR has been tested on two real-world datasets and has achieved better performance than common cross-domain sequential recommendation models. Rich ablation experiments also verify the effectiveness of LeCDSR's modules and the generated embeddings from the large model. Our implementation is available at this repository: https://github.com/solozhu/LeCDSR