Utilizing Users’ Negative Preferences in LLM-Based Recommendation System

Recommendation systems have significantly evolved with the integration of large language models (LLMs) which have attracted increasing attention due to their strong language understanding capabilities. In typical human-to-human recommendation processes, a recommender first presents a list of items and offers explanations based on the user’s preferences. The user then provides feedback allowing the recommenders to refine future recommendations with great accuracy. To implement this refinement process effectively in automated systems, it is crucial to understand both positive and negative user preferences. This dual perspective highlights the discrepancies between the user’s interests and recommended items. In this study, we propose an LLM-based recommendation method that emphasizes the selection effects of recommended items by analyzing users’ negative preferences. We evaluate the effectiveness of our approach through experiments using the LLaMA3.1 and MovieLens 1M dataset, comparing performance with existing baseline methods.