MASRank: Multi-Agent Collaborative Large Language Models for Zero-Shot Ranking in Recommender Systems : Advancing Beyond Single-Agent Limitations through Specialized Collaboration

Résumé fourni par la source

Large Language Models (LLMs) have demonstrated promising capabilities as zero-shot rankers in recommender systems, yet they suffer from inherent limitations including position bias, popularity bias, and difficulty in perceiving sequential order. This paper introduces MASRank, a novel Multi-Agent Collaborative framework that addresses these challenges through specialized agent cooperation. Our framework employs five specialized agents: a Sequence Understanding Agent for temporal pattern recognition, a Content Analysis Agent for multimodal feature processing, a Bias Detection Agent for systematic bias mitigation, a Personalization Agent for user-specific preference modeling, and a Coordinator Agent for intelligent result integration. Extensive experiments on three benchmark datasets (MovieLens-25M, Amazon-Electronics, Yelp-2023) demonstrate that MASRank achieves consistent improvements over singleagent LLM approaches, with up to 0.8% enhancement in NDCG@10 and 1.2% improvement in diversity metrics while maintaining computational efficiency through optimized agent coordination strategies.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.