Investigating Cross-Domain Capabilities of LLMs in Two-Stage Recommender Systems

Leveraging Large Language Models (LLMs) for recommendation is an emerging area with promising advancements across various use cases. Pursuing this line of research, we explore the complex and often overlooked cross-domain scenario in which user preferences from a source domain (e.g., Books) are used to recommend items in a target domain (e.g., Movies). We postulate that LLMs can significantly enhance cross-domain recommendations due to their pre-training on extensive web-scale corpora, which encompass diverse knowledge across languages and domains. However, two key limitations hinder the application of LLMs in this context: input size constraints and the computationally intensive autoregressive generation phase, making real-world implementation challenging. To tackle these issues, we revisit a recently proposed two-stage approach that employs a retrieval model to select a small list of the most relevant items from the catalog, which is subsequently refined by the LLM for final recommendations. This approach effectively addresses input size constraints and eliminates the need for a costly generation phase. Our experiments with various LLMs demonstrate their potential in the cross-domain recommendation context, while emphasizing the need for further research, as not all LLM variants consistently yield improvements.