CrossHGL: A Text-Free Foundation Model for Cross-Domain Heterogeneous Graph Learning

Xuanze Chen; Jiajun Zhou; Yadong Li; Shanqing Yu; Qi Xuan

cs.LGMar 29, 2026

CrossHGL: A Text-Free Foundation Model for Cross-Domain Heterogeneous Graph Learning

Authors: Xuanze Chen, Jiajun Zhou, Yadong Li, Shanqing Yu, Qi Xuan

Organizations: Xuanze Chen and Yadong Li are with the Institute of Cyberspace Security, Zhejiang University of Technology, Hangzhou 310023, China, and also with the Binjiang Cyberspace Security Institute of ZJUT, Hangzhou, 310056, China · Jiajun Zhou, Shanqing Yu and Qi Xuan are with the Institute of Cyberspace Security, Zhejiang University of Technology, Hangzhou 310023, China, with the Binjiang Cyberspace Security Institute of ZJUT, Hangzhou, 310056, China, and with the Soovar Technologies Co., Ltd., Hangzhou 310056, China

Abstract

Heterogeneous graph representation learning (HGRL) is essential for modeling complex systems with diverse node and edge types. However, most existing methods are limited to closed-world settings with shared schemas and feature spaces, hindering cross-domain generalization. While recent graph foundation models improve transferability, they often target homogeneous graphs, rely on domain-specific schemas, or require rich textual attributes. Consequently, text-free and few-shot cross-domain HGRL remains underexplored. To address this, we propose CrossHGL, a foundation framework that preserves and transfers multi-relational structural semantics without external textual supervision. Specifically, a semantic-preserving transformation strategy homogenizes heterogeneous graphs while encoding interaction semantics into edge features. Based on this, a prompt-aware multi-domain pre-training framework with a Tri-Prompt mechanism captures transferable knowledge across feature, edge, and structure perspectives via self-supervised contrastive learning. For target-domain adaptation, we develop a parameter-efficient fine-tuning strategy that freezes the pre-trained backbone and performs few-shot classification via prompt composition and prototypical learning. Experiments on node-level and graph-level tasks show that CrossHGL consistently outperforms state-of-the-art baselines, yielding average relative improvements of 25.1% and 7.6% in Micro-F1 for node and graph classification, respectively, while remaining competitive in challenging feature-degenerated settings.

Figures & tables

Fig. 1: Challenges in cross-domain heterogeneous pre-training.

Fig. 2: Overview of the CrossHGL framework. The workflow consists of three phases: 1) Semantic-preserving graph transformation : each source and target heterogeneous graph is converted into a unified homogeneous graph through SVD-based feature alignment and meta-pattern mining, while heterogeneous semantics are preserved in semantics-enriched edge features; 2) Multi-domain semantic graph pre-training : a Tri-Prompt mechanism with feature, edge, and structure prompts is integrated with a shared GNN backbone to capture transferable knowledge from feature, relation, and topology perspectives through self-supervised contrastive learning; 3) Cross-domain semantic fine-tuning : the pre-trained backbone is frozen, source-domain prompts are composed with target-specific open prompts for few-shot adaptation, and the adapted representations are used for prototype-based prediction on target-domain instances.

Dataset	#Nodes	#Node	#Edges	#Edge	#Classes
Dataset	#Nodes	Types	#Edges	Types	#Classes
ACM	10,942	4	547,872	8	3
DBLP	26,128	4	239,566	6	4
IMDB	11,616	3	34,212	4	3
Freebase	180,098	8	1,057,688	36	7

TABLE I: Summary of the heterogeneous graph datasets used in our cross-domain evaluation.

Paradigm	Category	Method	ACM		DBLP		IMDB		Freebase
Paradigm	Category	Method	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1
Supervised Learning	Heterogeneous	HAN	52.66 $\pm$ 3.90	49.20 $\pm$ 6.99	41.72 $\pm$ 5.52	36.68 $\pm$ 6.84	37.25 $\pm$ 2.15	27.68 $\pm$ 5.49	40.12 $\pm$ 7.39	8.01 $\pm$ 1.29
	Heterogeneous	Simple-HGN	48.15 $\pm$ 5.44	45.15 $\pm$ 3.67	37.53 $\pm$ 2.16	33.26 $\pm$ 4.16	35.14 $\pm$ 2.42	24.15 $\pm$ 4.31	37.73 $\pm$ 4.27	7.52 $\pm$ 2.32
	Homogeneous	GCN	61.02 $\pm$ 10.23	56.19 $\pm$ 12.92	37.95 $\pm$ 4.98	34.84 $\pm$ 6.16	37.09 $\pm$ 2.06	27.20 $\pm$ 4.92	22.42 $\pm$ 0.19	4.02 $\pm$ 0.21
		SAGE	47.83 $\pm$ 5.38	46.56 $\pm$ 6.10	44.01 $\pm$ 5.34	41.42 $\pm$ 5.82	37.67 $\pm$ 1.60	30.36 $\pm$ 4.15	29.21 $\pm$ 0.15	7.45 $\pm$ 0.03
		GAT	54.10 $\pm$ 6.96	46.75 $\pm$ 8.28	38.50 $\pm$ 4.47	33.91 $\pm$ 5.20	37.24 $\pm$ 2.05	31.49 $\pm$ 3.81	25.95 $\pm$ 0.14	5.30 $\pm$ 0.02

TABLE II: Few-shot node classification results: average Micro-F1 and Macro-F1 (%)

\pm

standard deviation across four cross-domain tasks. Boldface letters mark the best performance, while underlined letters indicate the second best.

Paradigm	Category	Method	ACM		DBLP		IMDB		Freebase
Paradigm	Category	Method	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1
Supervised Learning	Heterogeneous	HAN	47.47 $\pm$ 8.42	43.02 $\pm$ 10.98	49.13 $\pm$ 8.66	45.68 $\pm$ 9.38	34.20 $\pm$ 2.99	26.48 $\pm$ 6.38	34.13 $\pm$ 7.39	7.67 $\pm$ 2.48
	Heterogeneous	Simple-HGN	44.26 $\pm$ 5.16	40.02 $\pm$ 7.54	46.12 $\pm$ 6.26	43.15 $\pm$ 5.21	35.16 $\pm$ 3.15	27.34 $\pm$ 6.26	31.73 $\pm$ 4.27	5.32 $\pm$ 1.29
	Homogeneous	GCN	54.68 $\pm$ 9.08	48.26 $\pm$ 11.79	43.38 $\pm$ 6.28	37.64 $\pm$ 7.06	35.02 $\pm$ 2.85	27.36 $\pm$ 4.89	20.04 $\pm$ 0.41	9.40 $\pm$ 0.05
		SAGE	42.63 $\pm$ 5.59	36.44 $\pm$ 8.02	47.14 $\pm$ 8.32	43.64 $\pm$ 9.28	35.24 $\pm$ 2.52	28.80 $\pm$ 5.35	26.94 $\pm$ 4.41	9.87 $\pm$ 2.60
		GAT	42.61 $\pm$ 4.17	34.01 $\pm$ 5.57	44.67 $\pm$ 7.53	40.61 $\pm$ 8.80	34.89 $\pm$ 2.63	26.89 $\pm$ 5.65	21.19 $\pm$ 5.75	8.78 $\pm$ 0.94

TABLE III: Few-shot graph classification results. Average Micro-F1 and Macro-F1 (%)

\pm

standard deviation are reported. Boldface letters mark the best performance, while underlined letters indicate the second best.

Fig. 3: Impact of sample size ( $k$ ) on ACM classification performance

Fig. 4: Performance comparison under different pre-training source settings. The full multi-source setting consistently achieves the best overall performance, while reducing the number of source domains leads to clear performance degradation on several target datasets.

Panel of Table 8 — TABLE IV: Ablation study results of Tri-Prompt components on four target datasets. Boldface denotes the best performance.

Fig. 5: Comparison between direct flattening and our semantic-preserving graph transformation on GCN, GraphSAGE, and GAT under the 1-shot supervised setting, where the remaining nodes are split into validation and test sets following the protocol used in the main results.

Explore similar work

Empowering Heterogeneous Graph Foundation Models via Decoupled Relation Alignment
May 1, 2026Ziyu Zheng, Yaming Yang, Zhe Wang +2Heterogeneous GNNs Cross-Domain Transfer Learning
School of Computer Science and Technology, Xidian University, Xi’an, China 710071
CHoE: Cross-Domain Heterogeneous Graph Prompt Learning via Structure-Conditioned Experts
May 15, 2026Peiyuan Li, Yongqi Huang, Jitao Zhao +3Heterogeneous GNNs Few-Shot Learning
School of Computer Science and Technology, Tianjin University, Tianjin, China · Department of Health Technology and Informatics, and Department of Data Science and Artificial Intelligence, The Hong Kong Polytechnic University, Kowloon, Hong Kong
Handling Feature Heterogeneity with Learnable Graph Patches
Jun 16, 2026Yifei Sun, Yang Yang, Xiao Feng +4Graph Representation Learning Cross-Domain Transfer Learning
Zhejiang University Hangzhou, China · Huazhong University of Science and Technology Wuhan, China · Finvolution Group Shanghai, China

May 1, 2026J/K move · Enter open · S save

Empowering Heterogeneous Graph Foundation Models via Decoupled Relation Alignment

Ziyu Zheng, Yaming Yang, Zhe Wang, Ziyu Guan+1

School of Computer Science and Technology, Xidian University, Xi’an, China 710071

Heterogeneous GNNs Cross-Domain Transfer Learning Graph Foundation Models

Open paper AskPDF arXiv

While Graph Foundation Models (GFMs) have achieved remarkable success in homogeneous graphs, extending them to multi-domain heterogeneous graphs (MDHGs) remains a formidable challenge due to cross-type feature shifts and intra-domain relation gaps. Existing global feature alignment methods (PCA or SVD) enforce a shared feature space blindly, which distorts type-specific semantics and disrupts original topologies, inevitably leading to "Type Collapse" and "Relation Confusion". To address these fundamental limitations, we propose Decoupled relation Subspace Alignment (DRSA), a novel, plug-and-play relation-driven alignment framework. DRSA fundamentally shifts the paradigm by decoupling feature semantics from relation structures. Specifically, it introduces a dual-relation subspace projection mechanism to coordinate cross-type interactions within a shared low-rank relation subspace explicitly. Furthermore, a feature-structure decoupled representation is designed to decompose aligned features into a semantic projection component and a structural residual term, adaptively absorbing intra-domain variations. Optimized via a stable alternating minimization strategy based on Block Coordinate Descent, DRSA constructs a well-calibrated, structure-aware latent space. Extensive experiments on multiple real-world benchmark datasets demonstrate that DRSA can be seamlessly integrated as a universal preprocessing module, significantly and consistently enhancing the cross-domain and few-shot knowledge transfer capabilities of state-of-the-art GFMs. The code is available at: https://github.com/zhengziyu77/DSRA.

Abstract

Explore similar work

Empowering Heterogeneous Graph Foundation Models via Decoupled Relation Alignment

CHoE: Cross-Domain Heterogeneous Graph Prompt Learning via Structure-Conditioned Experts

Handling Feature Heterogeneity with Learnable Graph Patches