Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Efficient Hypergraph and Multi-LLM Agent Recommender Systems

Tendai Mukande, Esraa Ali, Annalina Caputo, Ruihai Dong, Noel OConnor

TL;DR

The paper tackles dynamic, multi-behavior recommender systems and the challenges of hallucination and high computational cost in generative LLM-based approaches. It introduces HGLMRec, a hypergraph-encoder-based recommender that leverages a Multi-LLM Agent (MoA) framework to fuse rich higher-order user-item interactions with prompts and refine recommendations through multiple specialized agents. Empirical results on three real-world datasets show that HGLMRec outperforms state-of-the-art baselines at lower computational cost, with ablations confirming the importance of the hypergraph encoder and MoA. The approach provides scalable, efficient personalized recommendations suitable for large-scale and resource-limited deployments, with future work focusing on fairness and reducing MoA latency.

Abstract

Recommender Systems (RSs) have become the cornerstone of various applications such as e-commerce and social media platforms. The evolution of RSs is paramount in the digital era, in which personalised 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 HGLMRec, a novel Multi-LLM agent-based RS that incorporates a hypergraph encoder designed to capture complex, multi-behaviour 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.

Towards Efficient Hypergraph and Multi-LLM Agent Recommender Systems

TL;DR

The paper tackles dynamic, multi-behavior recommender systems and the challenges of hallucination and high computational cost in generative LLM-based approaches. It introduces HGLMRec, a hypergraph-encoder-based recommender that leverages a Multi-LLM Agent (MoA) framework to fuse rich higher-order user-item interactions with prompts and refine recommendations through multiple specialized agents. Empirical results on three real-world datasets show that HGLMRec outperforms state-of-the-art baselines at lower computational cost, with ablations confirming the importance of the hypergraph encoder and MoA. The approach provides scalable, efficient personalized recommendations suitable for large-scale and resource-limited deployments, with future work focusing on fairness and reducing MoA latency.

Abstract

Recommender Systems (RSs) have become the cornerstone of various applications such as e-commerce and social media platforms. The evolution of RSs is paramount in the digital era, in which personalised 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 HGLMRec, a novel Multi-LLM agent-based RS that incorporates a hypergraph encoder designed to capture complex, multi-behaviour 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.

Paper Structure

This paper contains 20 sections, 6 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Graph and LLM Integration.
  4. Methodology
  5. Hypergraph Encoder
  6. Feature Initialisation.
  7. Hypergraph Convolution.
  8. Adaptive Readout.
  9. Token Generation.
  10. Token Fusion
  11. MoA Framework
  12. Model Training
  13. Experiments
  14. Hyperparameters.
  15. Datasets and Baseline Models.
  16. ...and 5 more sections

Figures (3)

  • Figure 1: An example of a real-world e-commerce recommendation scenario involving multiple user behaviour types. This scenario is best represented by a hypergraph, allowing modelling of higher-order, multi-way relationships. Nodes: users/items. Hyperedges: view, add-to-cart, purchase, etc
  • Figure 2: End-to-end architecture of HGLMRec. The model processes user-item interactions through an encoder, fuses graph tokens with task prompts, and refines recommendations via a 3-layer MoA framework. These agents leverage interactions captured by the encoder . The final MoA layer (L3,1) aggregates information from the intermediate LLM agents.
  • Figure 3: Cost and efficiency evaluation of the HGLMRec model on the IJCAI dataset. The cost (in US$) is calculated based on pricing information available on API provider websites.