Graph Foundation Models for Recommendation: A Comprehensive Survey

📝 Paper Summary

Graph Foundation Models (GFMs) Graph Neural Networks (GNNs) for Recommendation Large Language Models (LLMs) for Recommendation

This survey establishes a taxonomy for Graph Foundation Models in recommendation, categorizing methods by how they integrate the structural reasoning of GNNs with the semantic understanding of LLMs.

Core Problem

GNN-based recommenders struggle with textual semantics due to structural bias, while LLM-based recommenders lack the ability to model complex higher-order user-item interaction structures.

Why it matters:

Existing approaches utilizing only GNNs or LLMs fail to capture the complete picture of user preferences, which consists of both structural signals (interactions) and textual signals (profiles/descriptions)
Prior surveys focus on either GNNs or LLMs in isolation, or discuss Graph Foundation Models without addressing the specific challenges of the recommendation downstream task

Concrete Example: A GNN can recommend items bought by similar users but cannot understand why a user liked a specific item description. Conversely, an LLM can understand the description but misses the collaborative signal that 'users who bought A also bought B' if that path is not explicitly described in text.

Key Novelty

Taxonomy of Synergistic Graph-LLM Integration

Graph-augmented LLM: Uses the graph to provide structural context to the LLM, either by injecting graph tokens directly into the sequence or by converting graph paths into natural language prompts
LLM-augmented Graph: Uses the LLM's world knowledge to enhance the graph data itself, such as by predicting new edges (topology augmentation) or refining node features before feeding them into a GNN

Breakthrough Assessment

8/10

Provides a timely and necessary structuring of a rapidly emerging field (GFMs for RecSys), offering a clear taxonomy that organizes disparate recent works into a coherent framework.

⚙️ Technical Details

Problem Definition

Setting: Recommender Systems (RS) leveraging Graph Foundation Models (GFMs) to process both structural (user-item interactions) and textual (attributes/descriptions) information

Inputs: User-item interaction graphs, user profiles, item descriptions, and external knowledge graphs

Outputs: Taxonomy and categorization of methodologies for personalized recommendation ranking and generation

Comparison to Prior Work

vs. Gao et al. / Wu et al.: Integrates both fields (GNN + LLM) rather than treating them separately
vs. Liu et al.: Specifically addresses the application of GFMs to Recommender Systems, detailing taxonomies relevant to user-item interaction modeling

Limitations

Graph-augmented LLM methods rely heavily on the LLM, introducing inference latency and potential hallucinations
LLM-augmented Graph methods may be limited by the quality of the LLM's generated edges or features, potentially introducing noise into the graph structure
The survey notes that while GFMs integrate strengths, reasoning limitations of LLMs still restrict the depth of collaborative signals they can comprehend compared to pure GNNs

Reproducibility

This is a survey paper and does not provide a single codebase. It references numerous external papers (e.g., LLMGR, LLMRG, DALR) which may have their own repositories.

📚 Prerequisite Knowledge

Prerequisites

Understanding of Graph Neural Networks (GNNs) and their role in collaborative filtering
Familiarity with Large Language Models (LLMs) and prompting strategies
Basic concepts of Graph Foundation Models (pre-training on graph data)

Key Terms

GFM: Graph Foundation Model—a model pre-trained on large graph datasets that incorporates graph structures to solve graph-related tasks with generalization capabilities

Graph-augmented LLM: A category of methods where the LLM is the backbone, and structural information from graphs is injected (via tokens or context) to aid reasoning

LLM-augmented graph: A category of methods where the GNN is the backbone, and LLMs are used to enhance the graph's topology (edges) or features (text attributes)

Token-level infusion: Injecting graph information into the LLM by representing nodes or subgraphs as special tokens (e.g., [ACTION], [ITEM]) within the input sequence

Context-level infusion: Providing graph information to the LLM by converting graph structures into natural language descriptions or retrieving relevant subgraphs as context

Topology augmentation: Using LLMs to restructure the graph by predicting new edges (relationships) or adding nodes, thereby modifying the topological structure for downstream GNNs

Graph-LLM harmonization: Methods that align the semantic embedding space of LLMs and the structural embedding space of Graphs for mutual optimization