Understanding User Behavior in Cross-Domain Recommendation: An LLM-Based Approach

Dipak Falgun Meher, Ajay Krishna Vajjala, David S. Rosenblum

Abstract: Cross-domain recommendation (CDR) has emerged as a promising solution to address the cold-start and sparsity issues faced by single-domain recommender systems. Users often exhibit varying interests and rating behaviors across domains, such as rating items in the Movies domain differently than in the Books domain. In this work, we empirically analyze whether state-of-the-art CDR algorithms make significantly better recommendations in a target domain when a user’s rating behavior is consistent across domains. We propose a novel approach leveraging Large Language Models (LLMs) to quantify a consistency value that measures how consistently a user rates items across different domains. Our empirical analysis reveals that the performance of state-of-the-art CDR models does not consistently correlate with user behavior consistency across domain pairs, indicating limitations in their ability to effectively leverage this factor. These findings highlight the need for CDR algorithms that better utilize user behavior consistency to enhance recommendation performance.

External IDs:dblp:conf/ijcnn/MeherVR25