With the rapid development of Large Language Models (LLMs), their outstanding zero-shot text processing and generation capabilities, even without separate fine-tuning, are expanding their application scope. Consequently, LLMs have begun to be utilized...
With the rapid development of Large Language Models (LLMs), their outstanding zero-shot text processing and generation capabilities, even without separate fine-tuning, are expanding their application scope. Consequently, LLMs have begun to be utilized in recommendation systems, performing personalized recommendations and content ranking based on user profile information. However, most research on LLM-based recommendation systems operates as an augmenter (LLM-as-Augmenter) to traditional recommendation models, rather than as a recommender (LLM-as-Recommender). However, because the role of the LLM-as-Augmenter is performed in conjunction with existing recommendation models, the cold-start problem inherent in traditional recommendation models persists. In contrast, this paper focuses on the LLM-as-Recommender ranking recommendation system that overcomes the cold-start problem by leveraging its extensive inherent content understanding.
Research on utilizing LLMs as recommenders is still in its infancy, and concerns about hallucination and inherent implicit bias within the model itself are key issues in recommender systems. In particular, popularity bias in LLMs is problematic because they only recommend generally popular items, without considering users' personalized item preferences. This leads to a significant difference in recommendation quality compared to traditional recommendation models, depending on the detailed sensitive attributes (SST) of the user profiles included in the prompts. If uniform profile-based recommendations were made to all users regardless of their item preferences, this could result in lower recommendation quality for certain user groups. Therefore, in this paper, we predict user item preferences based on the item interaction history in the user profile and selectively incorporate user-SST information into the LLM prompts to mitigate the deterioration in recommendation quality caused by LLM popularity bias. The proposed method was validated through the LLM-as-Recommender ranking evaluation, demonstrating improved recommendation quality across various demographic groups.