GRAPHAM: A Graph-Powered Hierarchical Autonomous Multi-Agent System for Next-Generation Recommendations

Traditional recommender systems frequently en counter issues such as popularity bias, cold-start problems, and a lack of transparency. This paper presents GRAPHAM, a novel agentic framework that addresses these limitations by employing a society of collaborative AI agents. These agents, leveraging Large Language Models (LLMs) via the high-throughput Groq API, reason over a heterogeneous knowledge graph to generate recommendations. By orchestrating specialized agents for user profiling, diverse candidate generation, and meticulous ranking, GRAPHAM implements a sophisticated, human-like reasoning process without requiring model fine-tuning. We introduce a new agentic architecture that replaces a linear pipeline with a collaborative workspace, enhancing the system’s modularity and dynamism. A key contribution is the implementation of a batch-processing strategy in the ranking agent, which successfully overcomes LLM context window limits. We conduct a rigorous quantitative evaluation on the MovieLens dataset, comparing GRAPHAM’s performance against two deep learning baselines: the state-of-the-art LightGCN [8] and the classic Neural Collab orative Filtering (NCF) [5]. The results show that while graph based models excel in accuracy, GRAPHAM provides competitive performance with unparalleled explainability and a strong hit rate, demonstrating the viability of training-free, reasoning-based agentic systems for complex recommendation tasks.