PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Cross-domain Graph Learning: Progress and Future Directions

Haihong Zhao, Zhixun Li, Chenyi Zi, Aochuan Chen, Fugee Tsung, Jia Li, Jeffrey Xu Yu

arXiv:2503.11086v2cs.LGcs.AI

TL;DR

Graph learning struggles to generalize across domains because graph structures and features differ substantially. This survey organizes CDGL research into a taxonomy based on transferable knowledge, reviews representative methods, and identifies limitations and future directions. It positions CDGL as a route toward general-purpose graph foundation models.

  • Problem

    Graph learning lacks reliable cross-domain generalization because graph domains differ in structure, feature dimensionality, and feature semantics.

    Source: §I · p. 2, A. Feature Alignment · p. 13

  • Method

    The survey proposes a taxonomy of structure-oriented, feature-oriented, and mixture-oriented CDGL methods and systematically reviews representative studies.

    Source: §I · p. 2, §VIII · p. 15

  • Results

    The survey synthesizes current CDGL strategies, challenges, limitations, and future research directions across multiple cross-domain scales and difficulty levels.

    Source: §VIII · pp. 15–16

  • Takeaways & Limitations

    CDGL provides a framework for understanding progress toward graph models with transferable knowledge across diverse graph domains.

    Source: §VIII · p. 16

    Universal feature alignment remains unresolved because dimensional alignment does not guarantee semantic equivalence across domains.

    Source: A. Feature Alignment · pp. 13, 15

Abstract

from arXiv · show

Graph learning plays a vital role in mining and analyzing complex relationships within graph data and has been widely applied to real-world scenarios such as social, citation, and e-commerce networks. Foundation models in computer vision (CV) and natural language processing (NLP) have demonstrated remarkable cross-domain capabilities that are equally significant for graph data. However, existing graph learning approaches often struggle to generalize across domains. Motivated by recent advances in CV and NLP, cross-domain graph learning (CDGL) has gained renewed attention as a promising step toward realizing true graph foundation models. In this survey, we provide a comprehensive review and analysis of existing works on CDGL. We propose a new taxonomy that categorizes existing approaches according to the type of transferable knowledge learned across domains: structure-oriented, feature-oriented, and mixture-oriented. Based on this taxonomy, we systematically summarize representative methods in each category, discuss the key challenges and limitations of current studies, and outline promising directions for future research. A continuously updated collection of related works is available at: https://github.com/cshhzhao/Awesome-Cross-Domain-Graph-Learning.

I. INTRODUCTION

The introduction motivates cross-domain graph learning as a route toward graph foundation models and surveys the field through a structured taxonomy. It emphasizes that structural and feature differences make knowledge transfer difficult, while existing work lacks a comprehensive synthesis.

I. INTRODUCTION

  • Cross-domain graph learning seeks to transfer knowledge from diverse source graphs to target domains, supporting graph foundation models.The goal is to distill shared structural and feature patterns into domain-agnostic representations.Source: Figure 1 · p. 1, §I · pp. 1–2
  • Structural differences in connectivity and scale can trigger structure-level negative transfer across graph domains.Citation, collaboration, social, and molecular graphs exhibit markedly different connectivity patterns and graph sizes.Source: §I · p. 2
  • Feature differences in dimensionality and semantics hinder identifying commonalities and can prevent models from learning across domains.Citation features may be high-dimensional textual representations, whereas molecular features describe fewer atom properties.Source: §I · p. 2
  • Existing surveys mainly address domain adaptation or generalization under aligned feature spaces, leaving broader CDGL methods insufficiently integrated.Related methods often focus on narrower settings such as same high-level domains or distribution shifts.Source: §I · p. 2
  • This survey organizes CDGL research into structure-oriented, feature-oriented, and mixture-oriented approaches and reviews challenges and future directions.The taxonomy is presented as a structured overview of representative works.Source: §I · p. 2

B. Graph Representation Learning

Graph representation learning builds node, edge, or graph embeddings, while cross-domain learning seeks shared representations that generalize from source to target domains. Graphs remain harder to align than text or images because their structures and features vary substantially across domains.

B. Graph Representation Learning

  • Graph representation learning maps nodes, edges, or whole graphs into continuous spaces while preserving structural and feature information.The resulting representations support node classification, link prediction, and graph classification.Source: B. Graph Representation Learning · pp. 2–3
  • Shallow embeddings preserve network similarity but cannot represent unseen nodes or incorporate node features effectively.Node2vec and DeepWalk use factorization or random walks to construct low-dimensional embeddings.Source: B. Graph Representation Learning · p. 3
  • GNN message passing repeatedly aggregates neighbor information and updates central-node representations, enabling higher-order structural modeling.Stacking layers expands the receptive field to higher-order neighbors.Source: B. Graph Representation Learning · p. 3
  • Conventional graph representation methods are mainly designed for individual graph domains, limiting their effectiveness across domains.This motivates methods that generalize representations across diverse graph applications.Source: B. Graph Representation Learning · p. 3

C. Cross-domain Learning

  • Cross-domain learning learns a shared embedding space from multiple source domains and generalizes it to target-domain tasks.A source encoder maps inputs into the shared space, while task-specific functions map target embeddings to task results.Source: C. Cross-domain Learning · pp. 3–4
  • Unlike text and images, graph structures require manual construction and graph features vary widely or may lack semantic alignment, slowing cross-domain learning.These properties make shared representation formats harder to establish than in NLP and CV.Source: C. Cross-domain Learning · p. 4

D. Cross-domain graph learning

CDGL transfers knowledge from source graph domains to target graph domains for downstream tasks. The survey frames this process as an extension of cross-domain learning to graph-structured data.

D. Cross-domain graph learning

  • CDGL captures knowledge from multiple source graph domains and transfers it to task functions on target graph domains.Supported tasks include node classification and graph classification.Source: D. Cross-domain graph learning · p. 4

1) The Refinement of Cross-domain Scale:

The survey refines CDGL by distinguishing three scales based on domain compatibility and three difficulty levels based on source–target overlap. This framework clarifies how feature requirements and training overlap shape evaluation scenarios.

1) The Refinement of Cross-domain Scale:

  • Domain definitions determine CDGL scope, so the survey explicitly separates scale from difficulty when characterizing cross-domain settings.Within the same scale, difficulty varies with the intersection between source and target domains.Source: 1) The Refinement of Cross-domain Scale · pp. 4–5, 2) The Refinement of Cross-domain… · p. 5
  • CDGL spans limited, conditional, and open scales, depending respectively on shared attributes, transformable attributes within a high-level field, or arbitrary domains.Feature-dimension consistency typically indicates limited CDGL, shared representation mapping indicates conditional CDGL, and alignment of inherently diverse dimensions indicates open CDGL.Source: 1) The Refinement of Cross-domain Scale · p. 5

2) The Refinement of Cross-domain Difficulty:

  • High-difficulty CDGL has no source–target overlap, moderate difficulty has partial overlap, and low difficulty has identical source and target representation spaces.High-difficulty settings are associated with limited target data, while low-difficulty settings suit abundant-data scenarios using one model across domains.Source: 2) The Refinement of Cross-domain… · p. 5
  • Robust performance in high-difficulty settings is central to advancing open CDGL, although source–target overlap alone does not guarantee good generalization.Domain-adaptation methods using unlabeled target samples are treated as low difficulty because source and target are considered completely overlapping.Source: 2) The Refinement of Cross-domain… · p. 5

3) Proposed Taxonomy:

The survey organizes transferable graph knowledge into structure-, feature-, and mixture-oriented categories, then describes generative structure-oriented methods that unify graph connectivity across domains. These methods generate or reconstruct structures before transferring a model to target tasks.

3) Proposed Taxonomy:

  • The proposed taxonomy divides CDGL into structure-oriented, feature-oriented, and mixture-oriented approaches according to whether transferable knowledge comes from structure, features, or both.The taxonomy uses nodes, edges, adjacency matrices, and node features as the key graph elements.Source: 3) Proposed Taxonomy · p. 5

III. STRUCTURE-ORIENTED APPROACH

  • Structure-oriented approaches address cross-domain connectivity differences by extracting structural commonalities and dividing methods into generative and contrastive branches.The supplied subsection develops the generative branch; contrastive methods are summarized separately.Source: §III · pp. 5–6

A. Generative

  • Generative methods use a shared generator to create diverse structures for source and target graphs, giving models a unified structural perspective for target-domain generalization.The generated source structure is used to train a graph model, which is then transferred through a task-specific prediction head.Source: A. Generative · p. 6, Figure 3 · p. 6
  • GraphControl, UniAug, and GA2E transfer structural knowledge through feature-derived structures, diffusion-based modification, or subgraph reconstruction, respectively.These methods instantiate shared structural-pattern learning with different generation or reconstruction mechanisms.Source: A. Generative · p. 6
  • OpenGraph, GFT, and GFSE broaden structural transfer through unified node tokens, computation-tree vocabularies, or structure-aware self-supervised tasks.OpenGraph also uses LLM-synthesized node–edge sets to expand structural coverage, while GFSE models shortest-path substructures.Source: A. Generative · p. 6

B. Contrastive

Structure-oriented contrastive methods compare graph structures to identify domain-invariant commonalities. They create explicit or implicit positive–negative relationships, train graph models on source structures, and transfer them to target domains.

B. Contrastive

  • Contrastive methods compare or modify source graphs to construct positive and negative structural pairs before target-domain generalization.The contrastive module performs structural comparative analysis, while a task-specific prediction head produces target predictions.Source: B. Contrastive · pp. 6–7
  • GCC, PCRec, and APT use augmented or selected source graphs to form contrastive pairs that capture structural commonalities for downstream transfer.Their settings include generic subgraphs, user–item graphs without node features, and representative graphs selected by structural properties.Source: B. Contrastive · p. 7
  • GRADE measures source–target subtree differences with GSD and combines that term with task loss, implicitly pulling similar subtrees together and pushing dissimilar ones apart.The method is inspired by the Weisfeiler–Lehman test and uses the resulting structural signal for target generalization.Source: B. Contrastive · p. 7
  • FedStar, BooG, ProCom, and RiemannGFM capture shared structure through federated aggregation, virtual graph elements, community-level contrast, or geometric structural vocabularies.These approaches respectively emphasize multi-domain aggregation, unified structural characteristics, dual-level community learning, and Riemannian geometric contrast.Source: B. Contrastive · p. 7

C. Discussions

Structure-oriented methods handle complex graph connectivity and can generalize flexibly, but relying mainly on structure may overlook feature semantics. Feature-oriented methods therefore align semantic information in-space or feature dimensions across-space, with prompts optionally supporting transfer.

C. Discussions

  • Structure-oriented methods handle complex structural information and can generate meaningful structures for unseen domains, but may ignore coherence between generated structures and graph features.The survey identifies integrating structural and feature information as an important future direction.Source: C. Discussions · p. 7, D. Significance and Future Implications · p. 7

IV. FEATURE-ORIENTED APPROACH

  • Feature-oriented approaches learn transferable feature representations, separating in-space methods for semantic alignment from cross-space methods for dimension alignment.The distinction reflects whether source and target feature dimensions are already compatible.Source: §IV · pp. 7–8, Figure 4 · p. 8

A. In-space

  • In-space methods assume consistent feature dimensions and align feature semantics, optionally using prompts to make target features fit the pretrained semantic space.Their pipeline applies semantic alignment to source features, trains a graph model, and may prompt target features before task prediction.Source: A. In-space · p. 8
  • KTN, NaP, GPF-plus, RELIEF, DAGPrompT, and GADT3 implement in-space transfer through embedding transformations, contrastive relabeling, prompts, low-rank updates, or test-time training.These methods target semantic transfer while differing in whether they use source data, prompts, parameter updates, or target-specific self-supervision.Source: A. In-space · p. 8

IV. FEATURE-ORIENTED APPROACH

  • Cross-space approaches map features with mismatched dimensions into a shared embedding space, relaxing the scale limitations of in-space CDGL.The supplied figure contrasts semantic alignment with dimension alignment and notes that prompting is optional for both approaches.Source: Figure 4 · p. 8

B. Cross-space

Cross-space CDGL aligns graph features with different dimensions into a shared embedding space, allowing one graph model to learn across varied domains and generalize to targets.

B. Cross-space

  • Dimension alignment maps graph features with different dimensions into a shared embedding space for cross-domain graph learning.The aligned source features are then processed by a graph model trained for target-domain generalization.Source: B. Cross-space · pp. 8–9
  • A graph model trained on aligned source domains is generalized to target domains through a task-specific prediction head.A prompting module can modify target features to better match the representation space learned by the graph model.Source: B. Cross-space · pp. 8–9
  • LLM-based methods align cross-domain node features by generating shared embeddings, fine-tuning feature spaces, or adding semantic prompts for zero-shot prediction.OFA freezes an LLM for node embeddings, ZeroG applies LoRA and semantic prompting, and GraphAlign adds feature encoding, normalization, and mixture-of-experts alignment.Source: B. Cross-space · p. 9

C. Discussions

Feature-oriented methods address severe cross-domain feature mismatches, while mixture-oriented methods combine structural and feature knowledge through sequential or unified integration strategies.

C. Discussions

  • Feature-oriented methods can handle major source-target feature differences, but skipping dedicated structural extraction may create conflicts between structural and feature patterns.Such conflicts can limit the performance upper bound of cross-domain graph models.Source: C. Discussions · p. 9

D. Significance and Future Implications

V. MIXTURE-ORIENTED APPROACH

  • Mixture-oriented approaches jointly learn cross-domain structural and feature commonalities, and are classified as Sequential Mixture or Unified Mixture according to their fusion strategy.This direction explores structure and feature information either separately or in pairs.Source: §V · p. 9

A. Sequential Mixture

  • Sequential Mixture methods extract features and structures in succession, then use the resulting graph model for target-domain prediction.The extraction order can be reversed depending on the method design.Source: A. Sequential Mixture · p. 9
  • Structure-first sequential methods include SDGA, GSPT, SOGA, and MLDGG, which combine structural extraction with feature alignment or representation learning for cross-domain transfer.Their mechanisms include high-order structural extraction, unified feature spaces, separate structure-feature objectives, and disentanglement of invariant semantics.Source: A. Sequential Mixture · pp. 9–10
  • Figure 5 distinguishes sequential mixtures, which process feature and structure modules in order, from unified mixtures, which learn them jointly.The figure also includes GNN-based and flatten-based unified strategies for joint cross-domain knowledge extraction.Source: Figure 5 · p. 10
  • Feature-first sequential methods such as UniGraph and PTGB unify or align node features before using GNN-based learning to capture transferable structural commonalities.They subsequently transfer through in-context learning, graph instruction tuning, or target-domain fine-tuning.Source: A. Sequential Mixture · p. 10

B. Unified Mixture

Unified Mixture approaches jointly encode graph structure and features, either through GNN-based message passing or by flattening graphs into sequences processed by language models.

B. Unified Mixture

  • Unified Mixture methods simultaneously extract graph structures and features with a unified function before target-domain prediction.Their task heads typically consume representations produced by the unified graph model.Source: B. Unified Mixture · pp. 10–11
  • GNN-based unified methods differ mainly in message passing and information encoding while commonly using GNN-based task heads.Examples include domain adaptation, attention, hyperbolic representation, contrastive learning, optimal transport, and cross-domain edge construction.Source: B. Unified Mixture · p. 11
  • Flatten-based unified methods convert graphs into node or token sequences so LLMs can jointly process graph structure and semantics for downstream tasks.Representative strategies include generalized position embeddings, graph-to-token translators, instruction tuning, structured verbalizers, and GNN-LLM hybrids.Source: B. Unified Mixture · pp. 11–12
  • Pretraining-based unified methods use varied encoders, prompts, contrastive objectives, domain tokens, feature converters, or pivot nodes to align structural and feature information across domains.UniGraph2 additionally reconstructs node attributes and graph structures, while MDP-GNN aligns overall distributions with a Wasserstein-distance term.Source: B. Unified Mixture · p. 11
  • Other unified frameworks explore heterogeneous-graph domain generalization, meta-learning for brain connectomes, and federated learning across clients.These approaches aggregate heterogeneous contexts, adapt across self-supervised tasks, or cluster client models by parameter similarity.Source: B. Unified Mixture · p. 12

C. Discussions

Mixture-oriented methods integrate structural and feature information, but current feature alignment remains insufficient for broad open-domain scaling.

C. Discussions

  • Mixture-oriented methods more comprehensively capture cross-domain commonalities by integrating structural and feature information simultaneously.However, effective mapping of different-domain datasets into a shared feature space remains a major limitation.Source: C. Discussions · p. 12
  • Earlier mixture-oriented approaches scaled only to limited cross-domain settings, while newer LLM-based and projection-based methods have reached conditional or still exploratory open-domain scale.The cited discussion reports limited performance improvements from some domain-specific linear projection layers.Source: C. Discussions · p. 12

D. Significance and Future Implications

The survey frames CDGL around transferable structural, feature, and mixed knowledge, while highlighting open questions about topology–semantics alignment and transfer beyond originally graph-structured data. Future progress requires universal feature alignment across heterogeneous dimensions and semantics.

D. Significance and Future Implications

  • CDGL distinguishes structure-oriented, feature-oriented, and mixture-oriented transfer, with mixture methods seeking synergy between structural and feature representations.Mixture-oriented approaches must adapt inherently distinct structural and feature representation spaces to each other.Source: D. Significance and Future Implications · p. 12, §VI · p. 13

VI. OPEN QUESTIONS AND INSIGHTS

  • Shared motifs, hierarchical topologies, and community structures can provide domain-invariant relational principles that support structural transfer across heterogeneous graphs.Kernel-based similarity measures can normalize heterogeneous structures into comparable forms without using domain features.Source: §VI · p. 13
  • Unified transfer remains harder because features supply semantic meaning to topology, requiring models to align structural and feature spaces and discover their latent correspondence.LLM-based textual embeddings can help connect textual semantics with topological patterns through shared semantic representations.Source: §VI · p. 13
  • Open CDGL must determine whether graphs derived from NLP or CV data can transfer across domains, potentially extending graph-domain boundaries through cross-modal representations.Text-attributed graphs have transferred across textual domains, while visual graphs may leverage shared semantics through language grounding.Source: §VI · p. 13

A. Feature Alignment.

  • Universal feature alignment remains unresolved because heterogeneous domains differ in feature dimensions and semantics, and dimensional alignment alone may not ensure semantic equivalence.ProG, MDGPT [93], and CDFS-GAD [62] explore dimensional alignment with SVD or MLP, but open CDGL still needs a universal approach.Source: A. Feature Alignment · pp. 13, 15

B. Dealing with the Scale of Cross-domain Datasets.

The survey identifies dataset scale, data leakage, interpretability, evaluation, and domain compatibility as barriers to general-purpose CDGL. It concludes by organizing the field through a taxonomy and calling for systematic research directions toward graph foundation models.

B. Dealing with the Scale of Cross-domain Datasets.

  • Large-scale and diverse datasets are essential for improving open cross-domain performance, but traditional CDGL methods usually rely on smaller, limited or conditional data regimes.Structure-oriented methods can process larger graph data but omit feature information, while models using both structure and features remain constrained in dataset diversity and scale.Source: B. Dealing with the Scale of… · p. 15

C. Overcoming data leakage.

  • LLM-based CDGL evaluations risk data leakage because pretrained models may have memorized benchmark test data, motivating fairer and more comprehensive graph benchmarks.New post-cutoff citation datasets address only the citation domain and may not make graph structure sufficiently influential in evaluation.Source: C. Overcoming data leakage · p. 15

D. Improving interpretability.

  • Most CDGL methods remain difficult to interpret because GNN improvements are often based on intuition, while LLM explanations are largely limited to text-attributed graphs.Extending LLM reasoning and explanation from TAGs to non-TAGs matters for applications with strict security requirements.Source: D. Improving interpretability · p. 15

E. Unified Cross-domain Evaluation Metrics.

  • A unified cross-domain evaluation metric is needed because classical task metrics across datasets may not fully capture generalization from a global perspective.The proposed direction is to assess cross-domain generalization holistically rather than only through separate graph-task outcomes.Source: E. Unified Cross-domain Evaluation… · p. 15

F. Compatibility Analysis and Improvement across Domains.

  • Because information compatibility varies across domains, future CDGL should analyze compatibility and make targeted adjustments instead of treating all domains uniformly.The survey contrasts easier transfer between related fields with harder transfer between unrelated fields.Source: F. Compatibility Analysis and… · p. 15

VIII. CONCLUSION

  • The survey proposes a taxonomy of structure-, feature-, and mixture-oriented methods across limited, conditional, and open scales and high, moderate, and low difficulty levels.It systematically reviews representative studies and identifies remaining limitations and future research directions.Source: §VIII · pp. 15–16
Loading 2503.11086v2…