Graph Foundation Models for Recommendation: A Comprehensive Survey

• RS must integrate heterogeneous structural signals, such as user–item interactions, with textual information from profiles, descriptions, and reviews to improve relevance and robustness.Source: §1 · p. 2
• GNN-based RS capture higher-order collaborative dependencies but may underuse unstructured language, whereas LLM-based RS process text effectively yet can overlook collaborative filtering and interaction-graph structure [92] [100] [19].Source: §1 · pp. 2–3
• GFMs jointly model graph structures and language semantics by exchanging task-relevant signals between GNNs and LLMs [43].GNN-derived structural cues can be injected into LLM token-compatible forms, while LLM semantic priors can be aligned with graph representations for message passing.Source: §1 · p. 3
• GFM-based RS are presented as a promising paradigm for aligning user preferences with relational evidence and textual semantics, potentially yielding more accurate and robust recommendations.Source: §1 · p. 3

Taxonomy

• The survey taxonomy contains Graph-augmented LLM, LLM-augmented Graph, and LLM-Graph harmonization, representing injection into language models, graph enrichment before encoding, and shared-space alignment, respectively.Examples include LLMGR, LLMRG, and DALR.Source: §1 · p. 3

Survey contributions

• The survey fills a gap in comprehensive GFM-based RS coverage with a systematic overview of the field, taxonomy, methodological details, challenges, and future directions.It positions the work as a pioneering overview and reference for categorizing future research.Source: §1 · p. 3

2.1 Recommender Systems

• Sparse, biased interactions and strict serving constraints make representation learning central to generalizing across unseen items, cold-start entities, and evolving interests.Recommendation supervision is interaction-driven, while online retrieval and ranking must operate efficiently over large candidate spaces.Source: §2.1 · p. 4
• Interaction graphs capture higher-order collaborative dependencies, whereas natural-language content offers expressive semantics but does not automatically encode preference relevance.The survey therefore emphasizes combining both evidence types while keeping interaction-driven relevance as the optimization anchor.Source: §2.1 · p. 4

2.2 Graph Foundation Models

• The survey defines GFM-based recommender systems broadly as methods integrating graph-structured collaborative signals with foundation-model semantic capability, whether or not the graph component is itself pretrained.This RS-oriented definition includes systems built on pretrained components, commonly LLMs, and mechanisms that preserve or align interaction evidence with structural and semantic representations.Source: §2.2 · p. 5

2.3 GFM-based RS

• GFM-based recommendation is a design space in which graph propagation supplies collaborative structure, foundation models supply semantic generalization, and system design coordinates accuracy, robustness, and efficiency.The resulting representations and inference procedures must remain anchored to behavior-driven supervision and compatible with large-scale retrieval and ranking.Source: §2.3 · pp. 5–6, §2.2 · p. 5

2.4 Taxonomy of GFM-based RS

• The taxonomy groups methods into Graph-augmented LLM, LLM-augmented Graph, and LLM-Graph harmonization according to the dominant coupling direction between graphs and LLMs.Graph-augmented LLM methods inject structural knowledge into tokens or context; LLM-augmented Graph methods refine topology or node features; harmonization unifies semantic and structural spaces.Source: §2.4 · p. 6
• Graph-augmented LLM methods use token-level or context-level infusion to combine graph information with language-model reasoning and generation.Token-level methods represent nodes or subgraphs as special tokens, whereas context-level methods translate graphs into text or retrieve relevant text.Source: §2.4 · p. 6, Figure 3 · p. 7

3 Graph-Augmented LLM

• Graph-augmented LLM research addresses LLMs’ lack of an explicit mechanism for modeling user–item relationships and higher-order relational structure.The central design question is how to build cross-modal interfaces that bridge graph structures to language models.Source: §3 · p. 7

3.1 Token-Level Infusion

• Token-level infusion integrates nodes, subgraphs, actions, or graph embeddings as special input components, creating direct structural links to LLM processing and recommendation outputs.Examples include item tokens, syntax-integrated action tokens, graph-derived prefixes, and additive structural corrections such as e′_i = e_i + h_i.Source: §3.1 · p. 7
• Token-level infusion enables fine-grained interaction between structural and textual data but can require architectural modification or careful injection design.Parameter-efficient variants instead use learnable prefix and suffix tokens as adapters while preserving general model capabilities.Source: §3.1 · p. 8

3.2 Context-Level Infusion

• Context-level infusion supplies graph knowledge through serialized descriptions, retrieved records, or explicit reasoning chains without altering the LLM architecture.Methods verbalize graph paths, hypergraphs, multimodal relations, or domain constraints so the LLM can reason over them as natural-language context.Source: §3.2 · pp. 8–9
• Context retrieval methods increasingly filter or purify graph evidence because raw retrieved signals can introduce noise.Graph attention, collaborative paths, and voting ensembles are used to improve the fidelity of retrieved context, especially under sparse data or perturbations.Source: §3.2 · pp. 8–9
• Context-level systems also address cold start, latency, and interactive use through dynamic graph construction, early exits, and conversational retrieval of graph paths.These mechanisms balance recommendation performance with inference cost and support multi-turn dialogue and explanation generation.Source: §3.2 · p. 9

3.3 Discussion

• Graph-augmented LLMs can improve inference when interaction data is limited, but LLM reliance introduces risks including reduced diversity, distributional mismatch, scalability limits, and weaker generalizability.Alternative approaches mitigate these issues by enriching or harmonizing the graph, while retaining the LLM as a flexible plug-and-play component.Source: §3.3 · p. 9

4 LLM-Augmented Graph

• LLM-augmented graph methods divide into topology augmentation, which changes graph structure, and feature augmentation, which enhances textual or embedding information without changing topology.Both aim to improve GNN-based recommendation by augmenting graph data with LLM-derived information.Source: §4 · p. 9, Figure 4 · p. 10

4.1 Topology Augmentation

• Topology augmentation adds LLM-derived edges or nodes, enabling GNNs to learn richer user and item representations from structural priors and auxiliary information.Edge expansion captures complementary, substitutable, attribute-specific, inferred, or counterfactual relations, while node expansion adds interests, categories, intents, semantic codes, or other auxiliary concepts to compensate for limited interactions.Source: §4.1 · pp. 9–11

4.2 Feature Augmentation

• Feature augmentation improves graph node features without altering topology, including prompt-based enhancement of node textual information and GaCLLM’s textual message passing and aggregation within the GNN.Source: §4.2 · pp. 11–12

4.3 Discussion

• LLM-augmented graphs improve data-level recommendation capabilities in sparse cold-start scenarios, support plug-and-play model choices, and permit offline processing before online recommendation.However, plug-and-play graph methods may underuse LLM world knowledge and introduce extraneous noise.Source: §4.3 · p. 12
• Topology and feature augmentation may consequently introduce certain biases.Source: §4.3 · p. 12
• Their scalability is poor because the main methods use shallow GNNs whose depth is constrained by over-smoothing.Source: §4.3 · p. 12

5 LLM-Graph Harmonization

• LLM-graph harmonization addresses inefficiency, limited scalability, and weak knowledge utilization in graph-augmented LLMs and LLM-augmented graphs through a framework balancing efficiency, adaptability, and reasoning capabilities.Source: §5 · p. 12

5.1 Embedding Fusion

• Embedding fusion combines LLM-derived semantic embeddings with GNN-learned structural embeddings in a unified representation space using mechanisms such as concatenation or attention.This approach exploits complementary textual and relational information for recommendation representation learning.Source: §5.1 · pp. 12–13, Figure 5 · p. 13
• Dynamic fusion methods such as DynLLM adapt GNN-based interaction embeddings and LLM representations from real-time textual content before combining them in a shared latent space.DynLLM uses a dual-flow interaction mechanism to address static embeddings and support dynamic semantic and structural understanding.Source: §5.1 · p. 13
• Fusion-based systems integrate semantic and collaborative signals across applications, with LKPNR addressing long-tail news recommendation and broader fusion models improving robustness and expressiveness.The cited examples include high-quality news representations from LLMs combined with knowledge-graph relational structure.Source: §5.1 · p. 13

5.2 Embedding Alignment

• Embedding alignment maps LLM and graph embeddings into a shared space through contrastive learning, MLP-based transformations, or knowledge distillation to reduce heterogeneity and improve coherence.DALR aligns structural user-item graph representations with semantic product-description and review embeddings, while GLTA aligns graph nodes with LLM token embeddings.Source: §5.2 · p. 14
• Alignment methods support diverse recommendation settings: RLMRec improves learned representation quality, GLTA encodes structural relationships in the LLM vocabulary, and KEDRec-LM transfers knowledge-graph representations for medical recommendation.These methods provide coherent representations for personalized ranking, preference clustering, and knowledge-intensive recommendation tasks.Source: §5.2 · p. 14

5.3 Discussion

• Fusion directly combines modalities but may introduce redundancy, conflicts, and computational costs, whereas alignment mitigates inconsistencies but is sensitive to noise and high-quality training data.In both approaches, the LLM primarily functions as an encoder and cannot fully use its contextual understanding and language-generation capabilities under the described constraints.Source: §5.3 · p. 14

6 Datasets for GFM-based RS

• GFM-based RS datasets are grouped into Interaction-Centric, Semantic-Rich Textual, Knowledge-Driven Structured, and Conversational Dialogue categories, although many combine multiple signal sources.The taxonomy organizes datasets by the dominant data form in typical pipelines rather than treating the groups as mutually exclusive.Source: §6 · p. 16

6.1 Interaction-Centric Datasets

• Interaction-centric datasets provide large-scale behavior logs with limited text, testing graph-based modeling of sparsity, long-tail distributions, temporal dynamics, and higher-order connectivity.They appear as user–item bipartite graphs for long-term preferences or session graphs for short-term intent, making weak semantic channels a clean testbed for GFM structural gains.Source: §6.1 · p. 16
• Representative interaction benchmarks span explicit-rating datasets with minimal text, controlled rating-plus-tag settings, implicit timestamped events, and anonymous session sequences.Netflix emphasizes interaction topology, MovieLens permits varying semantic content, Last-FM supports sequential and popularity-bias analysis, while Diginetica and Yoochoose stress short-horizon intent and transition modeling.Source: §6.1 · pp. 16–17

6.2 Semantic-Rich Textual Datasets

• Semantic-rich textual datasets couple interaction graphs with reviews, descriptions, titles, or tags, enabling evaluation of semantic–collaborative alignment under noisy grounding, cold-start, and long-tail conditions.They also support studies of robust representation alignment, explanation- or reasoning-aware recommendation, and generalization when language does not directly encode preference strength.Source: §6.2 · p. 17
• Semantic-rich examples cover heterogeneous reviews and metadata, domain-specific multi-relational corpora, rapidly drifting news, long-form book narratives, multi-signal game behavior, and structured attributes.Yelp, Amazon, MIND, Goodreads, Steam, Book-Crossing, and Anime respectively expose challenges in review alignment, sparsity and domain shift, temporal generalization, textual abstraction, engagement-signal mismatch, cold-start metadata use, and attribute alignment.Source: §6.2 · pp. 17–18
• Semantic-rich textual datasets, including POI data, test whether models ground review opinions, aspects, and situational cues in user–place graphs despite geographic and popularity confounders.POI datasets include physical places, business categories, and sometimes coarse-grained location information.Source: §6.2 · p. 18

6.3 Knowledge-Driven Structured Datasets

• Knowledge-driven structured datasets provide explicit graphs, heterogeneous schemas, and concept hierarchies for relation-aware message passing, multi-hop reasoning, and semantic grounding.DBbook supports knowledge-grounded cold-start studies under sparse interactions, while DBLP and OpenAlex test heterogeneous, multi-relational scholarly reasoning and scalability.Source: §6.3 · pp. 18–19
• Junyi and ASSISTments evaluate whether graph–language models align concept structure with semantic descriptors while representing temporal learning dynamics and evolving mastery.These education datasets connect students, problems, and concepts through sequential interactions, knowledge components, or skill tags.Source: §6.3 · p. 19

6.4 Conversational Dialogue Datasets

• Conversational dialogue datasets evaluate unified dialogue understanding, preference elicitation, and graph-based item grounding under evolving user context.ReDial emphasizes coherent movie recommendations from dialogue-based preference revelation, whereas INSPIRED adds social strategies and socially appropriate response generation.Source: §6.4 · p. 19

Overview

• GFMs combine graph structural information with LLM-based representations, but broader adoption is hindered by unresolved effectiveness and deployment challenges.The survey presents these issues as central directions for advancing GFM-based recommendation.Source: §7 · p. 20

Knowledge-Preference Gap

• LLM-generated recommendations can be factually coherent yet insufficiently aligned with users’ preferred styles or interests.The survey motivates preference-aware knowledge adaptation, dynamic refinement, and user-specific contrastive learning to address this mismatch.Source: §7 · pp. 20–21

High Computational Cost and Scalability Issues

• High memory use and slow inference constrain GFM deployment on dense graphs and real-time personalized recommendation.The survey attributes these constraints to long-range graph dependencies and LLM-generated representations, motivating compression and adaptive graph sparsification.Source: §7 · p. 20

Robustness Against Noisy and Adversarial Data

• Robustness remains difficult because noisy, incomplete, biased, manipulated, and adversarial signals arise from both structured and unstructured data.Suggested directions include self-supervised denoising, multimodal adversarial training, and uncertainty-aware modeling.Source: §7 · p. 20

Multi-Modal Information Fusion

• Integrating text, graphs, images, audio, and video is challenging because modalities differ in granularity, semantic alignment, and computational cost.The survey points to adaptive fusion, cross-modal alignment, and lightweight multimodal representation learning.Source: §7 · p. 20

Lack of End-to-End Optimization

• End-to-end recommendation could reduce the time and manpower required by multi-stage matching, ranking, and re-ranking optimization.Generative recommendation combines multiple stages, and integrated structural-textual modeling is identified as a promising direction [82].Source: §7 · p. 20
• Generative recommendation that combines matching and ranking is described as a likely future research hotspot.The survey notes that such integration can reduce complexity and potentially improve performance, with industrial efforts already emerging [100].Source: §7 · p. 20

Knowledge-Preference Gap

• A knowledge-preference gap persists between globally pretrained world knowledge and personalized user preferences [73].This differs from embedding alignment because it concerns how LLMs interpret knowledge versus how users express preferences.Source: §7 · p. 20

Conclusion

• Graph foundation models are positioned as likely catalysts for a new wave of research in recommender systems.Source: §8 · p. 21
• The survey presents what it describes as the first comprehensive overview of GFM-based recommender systems.Source: §8 · p. 21
• It organizes the field through a logical taxonomy and examines its challenges and future potential.Source: §8 · p. 21