Graph Foundation Models for Recommendation: A Comprehensive Survey

TL;DR

GFMs for recommender systems fuse GNN-based relational modeling with LLM-based textual understanding to leverage both graph structure and natural language signals. The paper delivers a clear taxonomy—Graph-augmented LLM, LLM-augmented graph, and graph-LLM harmonization—and surveys concrete methods across token-level and context-level infusion, topology and feature augmentation, and embedding fusion/alignment. It identifies key challenges, including computational cost, robustness to noisy data, multi-modal fusion, end-to-end optimization, and knowledge-personalization gaps, and outlines directions to address them. This survey provides a structured reference for researchers and practitioners aiming to deploy GFMs in large-scale, real-world recommender systems, especially in cold-start and multi-modal contexts.

Abstract

Recommender systems (RS) serve as a fundamental tool for navigating the vast expanse of online information, with deep learning advancements playing an increasingly important role in improving ranking accuracy. Among these, graph neural networks (GNNs) excel at extracting higher-order structural information, while large language models (LLMs) are designed to process and comprehend natural language, making both approaches highly effective and widely adopted. Recent research has focused on graph foundation models (GFMs), which integrate the strengths of GNNs and LLMs to model complex RS problems more efficiently by leveraging the graph-based structure of user-item relationships alongside textual understanding. In this survey, we provide a comprehensive overview of GFM-based RS technologies by introducing a clear taxonomy of current approaches, diving into methodological details, and highlighting key challenges and future directions. By synthesizing recent advancements, we aim to offer valuable insights into the evolving landscape of GFM-based recommender systems.

Figures (5)

• Figure 1: An overview of GFM-based RS. Compared with GNN-based or LLM-based RS, GFM-based RS are positioned as integrating both approaches to create more comprehensive recommendations.
• Figure 2: A taxonomy of GFM-based recommender systems.
• Figure 3: The illustration of graph-augmented LLM methods: a) Token-Level Infusion, where nodes or subgraphs are represented as special tokens, integrating into LLM's input. b) Context-Level Infusion, where graph information is converted into context by translating graph into text or retrieving relevant text.
• Figure 4: The illustration of LLM-augmented graph methods: a) Topology Augmentation, where LLMs extract structural information from data to alter and augment the topological structure of the graphs.; b) Feature Augmentation, where LLMs processe the textual information in the data, augment the node text or embedding in the graph without changing the topological structure.
• Figure 5: The illustration of LLM-graph harmonization methods: a) Embedding Fusion, where LLM-derived semantic embeddings and GNN-learned structural embeddings are combined into a unified representation space through fusion mechanisms such as concatenation or attention-based integration; b) Embedding Alignment, where embeddings from both modalities are mapped into a shared space using techniques like contrastive learning or MLP-based transformation to enhance consistency and coherence.