Towards Efficient Hypergraph and Multi-LLM Agent Recommender Systems

Towards Efficient Hypergraph and Multi-LLM Agent Recommender Systems Tendai Mukande Research Ireland ML-LABS Dublin City University Dublin, Ireland tendai.mukande2@mail.dcu.ie Esraa Ali ADAPT Centre Dublin City University Dublin, Ireland abdelmoe@tcd.ie Annalina Caputo School of Computing Dublin City University Dublin, Ireland annalina.caputo@dcu.ie Ruihai Dong Insight Research Ireland Centre for Data Analytics University College Dublin Dublin, Ireland ruihai.dong@ucd.ie Noel O’Connor Insight Research Ireland Centre for Data Analytics Dublin City University Dublin, Ireland Noel.OConnor@dcu.ie Abstract Recommender Systems (RSs) have become the cornerstone of vari- ous applications such as e-commerce and social media platforms. The evolution of RSs is paramount in the digital era, in which per- sonalised user experience is tailored to the user’s preferences. Large Language Models (LLMs) have sparked a new paradigm - generative retrieval and recommendation. Despite their potential, generative RS methods face issues such as hallucination, which degrades the recommendation performance, and high computational cost in practical scenarios. To address these issues, we introduce HGLM- Rec, a novel multi-LLM agent-based RS model that incorporates a hypergraph encoder designed to capture complex relationships between users and items. The HGLMRec model retrieves only the relevant tokens during inference, reducing computational overhead while enriching the retrieval context. Experimental results show performance improvement by HGLMRec against state-of-the-art baselines at lower computational cost. CCS Concepts • Information systems →Recommender systems. Keywords LLM, Mixture of Agents, Hypergraph Neural Networks, Computa- tional Efficiency. ACM Reference Format: Tendai Mukande, Esraa Ali, Annalina Caputo, Ruihai Dong, and Noel O’Connor. 2025. Towards Efficient Hypergraph and Multi-LLM Agent Rec- ommender Systems. In Proceedings of (Preprint). ACM, New York, NY, USA, 8 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn This work is licensed under a Creative Commons Attribution 4.0 International License. Preprint, © 2025 Copyright held by the owner/author(s). ACM ISBN 978-x-xxxx-xxxx-x/YYYY/MM https://doi.org/10.1145/nnnnnnn.nnnnnnn 1 Introduction In real-world recommendation scenarios, user preferences evolve continuously over time, leading to complex and dynamic interaction patterns [23]. Static representation learning models [9, 16, 31] often struggle to capture these dynamics as they assume a static structure and ignore the time-dependent nature of user-item interactions [44]. With the recent surge in LLMs, a new paradigm in RSs has emerged that combines information retrieval with LLMs to produce contextually relevant recommendations [5, 20, 26, 29]. Generative recommendation models have shown benefits, such as semantic understanding and interactive reasoning, which can improve the relevance and quality of recommendations by generating output that aligns with user preferences [21, 33]. Despite these advances, most of the existing LLM-based RS ap- proaches face two major limitations. Firstly, hallucination, where the model generates inaccurate or misleading recommendations, can compromise the reliability of the system [13–15]. Secondly, the high computational cost, resulting from the need to search through large vocabularies or fine-tune LLMs on domain-specific data, makes these methods impractical for real-time or large-scale deployment [6, 34, 41]. Although pretraining or fine-tuning LLMs in recommendation-specific datasets can improve performance, these strategies require substantial computational resources [11, 22, 42], domain expertise, and large volumes of high-quality data, further complicating real-world implementation [46]. Consequently, effi- cient recommendation models are needed that can adapt to evolving user preferences and dynamic interaction patterns while maintain- ing high accuracy [12, 36]. Motivated by these challenges, we explore whether hypergraph representation learning can be harnessed to improve recommen- dation performance in dynamic, multi-behaviour scenarios. As il- lustrated in Figure 1, hypergraphs allow modelling of higher-order user-item interactions, unlike bipartite graphs that are limited to pairwise interactions [19]. We propose HGLMRec, a novel frame- work that integrates an HGNN encoder with an MoA architecture. The central idea of HGLMRec is to enhance modelling of user-item arXiv:2512.06590v1 [cs.IR] 6 Dec 2025 Preprint, Tendai Mukande, Esraa Ali, Annalina Caputo, Ruihai Dong, and Noel O’Connor interactions, allowing the model to capture higher-order depen- dencies across multiple behaviours. Hyperedges in this represen- tation connect a user with multiple items and behavioural types, generating dense token embeddings that encode local and global preference patterns [2]. These embeddings are then processed by the MoA framework, which employs multiple specialised agents to refine rec

This content is AI-processed based on open access ArXiv data.