[1H3-OS-40] OS-40

2:10 PM - 2:25 PM JST(5:10 AM - 5:25 AM UTC)

[1H3-OS-40-03] Enhancing Sequential Recommendation Using LLM-Generated Tags with a Multi-Agent Judging Framework

〇Zifan Zhang¹, Junichiro Mori¹ (1. The University of Tokyo)

[[online]]

Keywords:

Information recommendation,LLM,LLM-as-a-judge

Sequential recommendation models predict users’ next actions from interaction sequences, yet they often rely on item IDs or sparse metadata, limiting semantic generalization.We propose an LLM-enhanced framework that generates semantic tags for each item and refines them via LLMs-as-Judges.Our judging module supports both a single-judge baseline and a multi-agent setting in which specialized judges evaluate tag quality from complementary perspectives (accuracy, relevance, diversity, and usefulness) and aggregate their rankings to select high-quality tags.The selected tags are encoded by a pretrained BERT encoder and injected into a GRU-based sequential recommender.Experiments on MovieLens 20M show that LLM-generated tags outperform plot-based features and that judging-based filtering further improves next-item recommendation performance.