ReCoG: Relative Node Contexts for Multi-Domain Graph Pre-training

Yupeng Tong; Lianze Shan; Jitao Zhao; Dongxiao He

cs.LGNov 1, 2026

ReCoG: Relative Node Contexts for Multi-Domain Graph Pre-training

Authors: Yupeng Tong, Lianze Shan, Jitao Zhao, Dongxiao He

Organizations: School of Computer Science and Technology, Tianjin University, Tianjin, China

Abstract

The success of foundation models in language and vision has motivated Graph Foundation Models (GFMs), which learn transferable knowledge through multi-domain graph pre-training. Existing methods advance GFMs through feature unification and structural learning, and in doing so they rely on the raw graph structure to organize node contexts. However, we observe that removing a large proportion of edges from source graphs during pre-training improves cross-domain transfer accuracy across four representative methods. Our theoretical analysis identifies the raw graph structure as a source of conflict in cross-graph model sharing. We therefore advocate defining node contexts using criteria consistent across graphs. Achieving this requires (1) constructing node contexts consistently across graphs without a shared positional structure, such as the token order of text or the pixel grid of images, and (2) retaining the structural and neighborhood information of the original graphs. To this end, we introduce Relative Contexts for Graphs (ReCoG) for multi-domain graph pre-training. Specifically, we first construct attribute contexts to provide a comparison rule independent of the raw graph structure. Furthermore, we construct structural contexts based on common neighbors and semantic contexts based on neighborhood attribute summaries to retain the topological and neighborhood information of the original graphs. Since all graphs share these construction rules, we pre-train a single graph encoder by reconstructing node attributes over these contexts, without domain-specific modules. We conduct extensive experiments on 18 target datasets, which demonstrate the effectiveness of ReCoG. In particular, ReCoG exceeds the strongest baseline by 7.61 percentage points in mean 1-shot cross-domain node classification accuracy and by 4.47 points in cross-type transfer to heterophilous graphs.

Figures & tables

Figure 1. Mean 1-shot accuracy gains (percentage points) across six unseen target graphs under source-edge removal, relative to the same model pre-trained on full source graphs. Line chart of mean one-shot accuracy gains for MDGPT, TIG, MDGFM, and BRIDGE at source-edge removal ratios of zero, thirty, fifty, seventy, and ninety percent. MDGPT and TIG rise across the tested ratios, while MDGFM and BRIDGE peak at seventy percent. All four have positive gains at each nonzero removal ratio.

Figure 2. Recorded relations determine node contexts. (a) The same facts expressed as text and labeled graph edges. (b) Communication and task dependencies give the same agents different neighbors. Comparison of context relations in text, a knowledge graph, and two graphs of language-model agents with different recorded relations.

Figure 3. Overview of ReCoG. (a) Shared criteria construct and symmetrize attribute, structural, and semantic contexts. (b) A shared feature projector (MLP) and context aggregator (GNN) reconstruct attributes using a deduplicated context union. (c) Context weights

s_{r}=\lambda b_{r}

configure target propagation; the frozen encoder supports node classification, link prediction, and graph classification through task-specific predictors. The layered diagram uses blue, ochre and green for attribute, structural and semantic contexts. Projection guides join corresponding plane corners and fade beneath translucent surfaces; thin side faces and contact shadows indicate depth. Source graphs are visited in turn and share feature projection, context aggregation and attribute decoding. A distinct seven-node target graph uses configurable context weights and the frozen encoder. Three readout rows show classification of a selected node, prediction of a candidate link, and classification of a whole graph after pooling.

Method	Cora	CiteSeer	PubMed	CS	Photo	Computers	Avg.
FUG	46.40 $\pm$ 8.37	36.07 $\pm$ 6.81	52.23 $\pm$ 9.47	78.02 $\pm$ 5.88	63.62 $\pm$ 9.77	51.27 $\pm$ 12.82	54.60
TIG	48.88 $\pm$ 8.80	44.71 $\pm$ 9.02	50.94 $\pm$ 10.00	75.28 $\pm$ 6.22	63.52 $\pm$ 10.47	49.47 $\pm$ 13.08	55.47
GFT	36.19 $\pm$ 5.39	32.51 $\pm$ 5.36	50.56 $\pm$ 7.19	63.49 $\pm$ 6.42	49.91 $\pm$ 7.11	38.67 $\pm$ 7.91	45.22
MDGPT	42.49 $\pm$ 7.94	45.13 $\pm$ 8.81	55.34 $\pm$ 8.32	72.98 $\pm$ 6.61	64.19 $\pm$ 9.92	49.24 $\pm$ 10.88	54.90
SAMGPT	44.81 $\pm$ 8.42	41.52 $\pm$ 8.15	46.20 $\pm$ 9.67	76.48 $\pm$ 6.17	59.17 $\pm$ 8.63	46.73 $\pm$ 9.64	52.49
MDGFM	42.43 $\pm$ 6.05	38.84 $\pm$ 7.54	49.73 $\pm$ 8.60	74.11 $\pm$ 6.55	64.24 $\pm$ 9.52	49.63 $\pm$ 9.83	53.16

Table 1. Q1 1-shot node classification accuracy (%, mean

\pm

SD) over 500 shared tasks per held-out graph. Avg. denotes the six-target mean. Bold and underlined values mark the best and runner-up.

Figure 4. Q1 few-shot node classification accuracy (%) across 1-, 3-, and 5-shot budgets on six held-out target graphs (500 shared tasks per budget). ReCoG is compared against five representative baselines; full results appear in Tables 1 and 6 . Six panels showing 1-, 3-, and 5-shot node classification accuracy on Cora, CiteSeer, CS, PubMed, Photo, and Computers for FUG, TIG, MDGFM, BRIDGE, GraphLoRA, and ReCoG.

Method	Cornell	Texas	Wisconsin	Actor	Chameleon	Squirrel	Avg.
FUG	48.92 $\pm$ 8.84	62.70 $\pm$ 4.80	56.47 $\pm$ 3.49	28.30 $\pm$ 1.37	35.92 $\pm$ 2.81	26.77 $\pm$ 1.28	43.18
TIG	47.03 $\pm$ 7.57	58.92 $\pm$ 3.59	50.20 $\pm$ 5.13	27.52 $\pm$ 1.14	36.05 $\pm$ 1.70	26.12 $\pm$ 0.82	40.97
GFT	71.62 $\pm$ 6.65	78.11 $\pm$ 4.26	83.53 $\pm$ 4.49	33.57 $\pm$ 0.79	45.07 $\pm$ 1.89	35.06 $\pm$ 1.02	57.83
MDGPT	61.08 $\pm$ 8.13	62.97 $\pm$ 2.97	55.29 $\pm$ 3.37	28.44 $\pm$ 0.86	40.86 $\pm$ 1.87	33.13 $\pm$ 1.35	46.96
SAMGPT	60.54 $\pm$ 6.07	62.43 $\pm$ 4.43	56.86 $\pm$ 4.21	28.86 $\pm$ 0.88	42.11 $\pm$ 3.34	32.98 $\pm$ 0.99	47.30
MDGFM	61.62 $\pm$ 7.72	62.97 $\pm$ 5.14	55.88 $\pm$ 7.14	28.96 $\pm$ 1.25	42.65 $\pm$ 2.75	32.90 $\pm$ 0.97	47.50

Table 2. Q2 cross-type transfer accuracy (%, mean

\pm

SD) over ten splits. Avg. denotes the six-target mean. Bold and underlined values mark the best and runner-up.

	Node classification (Acc.)				Link prediction (AUC)				Graph classification (Acc.)
Method	Cora	CiteS.	PubM.	Avg.	Phys.	Wiki.	DBLP	Avg.	IMDB	COLL.	PROT.	Avg.
FUG	66.15 ±4.59	58.84 ±3.59	64.97 ±5.44	63.32	98.02 ±0.09	98.74 ±0.07	92.50 ±0.13	96.42	68.26 ±2.16	75.76 ±0.96	71.14 ±3.24	71.72
TIG	66.53 ±3.94	61.04 ±3.25	63.24 ±4.68	63.60	98.54 ±0.05	91.32 ±0.35	93.73 ±0.18	94.53	69.64 ±1.84	76.30 ±1.36	72.01 ±2.23	72.65
GFT	38.76 ±3.54	37.23 ±2.85	54.09 ±4.71	43.36	98.02 ±0.04	98.54 ±0.03	87.13 ±0.39	94.56	70.05 ±2.84	75.59 ±1.26	72.42 ±1.63	72.69
MDGPT	60.03 ±3.79	59.74 ±3.37	64.09 ±4.53	61.29	98.24 ±0.03	99.01 ±0.04	92.25 ±0.39	96.50	69.32 ±1.58	73.30 ±1.27	71.82 ±1.96	71.48
SAMGPT	59.80 ±4.36	59.44 ±3.53	56.41 ±7.00	58.55	98.25 ±0.04	98.95 ±0.06	91.14 ±0.30	96.11	70.77 ±1.52	74.29 ±1.06	72.41 ±1.16	72.49

Table 3. Q3 transfer from ogbn-products: 5-shot node classification, link prediction, and graph classification. Entries are accuracy or AUC (%, mean ±SD ); each Avg. weights three targets equally. Bold and underlined values mark the best and runner-up.

Figure 5. Q2 accuracy (%) by context type and strength

\lambda

. Dashed: self-only encoding (

\lambda=0

). Six panels showing how individual context types respond to propagation strength on Cornell, Texas, Wisconsin, Actor, Chameleon, and Squirrel.

Figure 6. Test-time robustness: (a) Gaussian feature noise, (b) edge removal, and (c) edge insertion. Mean accuracy (%) over the six Q1 graphs, with rebuilt contexts and refitted classifiers on frozen representations. Three line-plot panels comparing six-graph mean accuracy of six methods under feature noise, edge removal, and edge insertion.

Variant	Cora	CiteS.	PubM.	CS	Photo	Comp.	$\Delta$ Avg.
Full ReCoG	56.47 ±10.04	53.62 ±10.14	54.15 ±9.58	84.40 ±5.84	73.25 ±9.03	56.59 ±10.37	–
w/o Attribute	58.35 ±10.14	53.10 ±10.17	54.46 ±9.90	83.48 ±6.23	70.68 ±9.22	54.73 ±10.29	$-0.61$
w/o Structural	52.66 ±9.80	52.33 ±9.98	52.03 ±9.07	83.17 ±6.18	71.40 ±9.05	55.14 ±10.17	$-1.96$
w/o Semantic	49.74 ±9.62	47.63 ±9.42	53.95 ±9.89	83.23 ±6.06	72.15 ±8.67	55.89 ±10.56	$-2.65$
Raw edges	34.95 ±7.39	36.44 ±7.43	49.84 ±8.58	79.72 ±5.75	65.77 ±9.96	49.90 ±10.46	$-10.31$
Contrastive	48.57 ±8.27	30.59 ±5.14	41.21 ±6.37	77.48 ±6.59	68.11 ±8.33	56.79 ±10.27	$-9.29$

Table 4. Q1 1-shot ablation accuracy (%, mean ±SD ) over 500 shared tasks per target.

\Delta

Avg. is the six-target mean change (pp) from full ReCoG (bold; 63.08% mean).

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Nodes	Edges	Feat.	Cls.
Cora	2,708	5,278	1,433	7
CiteSeer	3,327	4,552	3,703	6
PubMed	19,717	44,324	500	3
CS	18,333	81,894	6,805	15
Photo	7,650	119,081	745	8
Computers	13,752	245,861	767	10

Appendix

Table 5. Dataset statistics. Labels of link-prediction targets are unused (–); IMDB-B and COLLAB use one-hot node degrees (deg.).

	Cora		CiteSeer		PubMed		CS		Photo		Computers		Avg.
Method	3	5	3	5	3	5	3	5	3	5	3	5	3	5
DGI	45.56 ±5.08	53.07 ±3.79	44.96 ±4.42	50.55 ±3.54	55.58 ±5.12	60.34 ±4.47	85.26 ±2.08	87.77 ±1.20	70.17 ±5.23	74.10 ±4.33	56.98 ±7.27	63.16 ±4.49	59.75	64.83
GRACE	38.59 ±4.00	44.50 ±3.25	30.63 ±3.88	34.72 ±3.42	45.68 ±4.85	49.95 ±4.23	65.49 ±3.51	71.96 ±2.75	57.22 ±6.52	64.24 ±5.42	50.64 ±6.08	56.71 ±5.15	48.04	53.68
BGRL	46.74 ±4.09	53.28 ±3.37	31.66 ±3.58	35.74 ±3.11	51.06 ±5.51	56.07 ±4.68	75.12 ±3.29	81.12 ±2.22	72.63 ±7.23	77.83 ±5.18	62.05 ±8.43	67.13 ±6.08	56.54	61.86
FUG	61.63 ±5.42	67.31 ±3.23	47.62 ±4.94	53.28 ±3.86	60.82 ±6.32	64.51 ±4.64	86.55 ±1.70	88.13 ±1.13	65.91 ±8.90	66.29 ±7.50	58.85 ±10.17	62.83 ±8.18	63.56	67.06
TIG	58.53 ±6.17	63.70 ±4.59	56.39 ±5.03	60.66 ±3.45	56.48 ±7.81	59.64 ±6.58	83.63 ±3.04	86.62 ±1.98	68.24 ±10.30	69.20 ±8.48	56.31 ±11.03	60.80 ±8.49	63.26	66.77

Appendix

Table 6. Q1 accuracy (%, mean ±SD ) at 3 and 5 shots over 500 shared tasks per target. Avg. weights six targets equally; 1-shot results appear in Table 1 . Bold and underlined values mark the best and runner-up.

Method	Cora	CiteSeer	PubMed	CS	Photo	Computers	Avg.
DGI	75.47 ±0.78	65.76 ±1.19	82.83 ±0.24	92.82 ±0.18	91.48 ±0.26	87.24 ±0.21	82.60
GRACE	72.00 ±0.36	56.29 ±0.71	76.80 ±0.37	90.29 ±0.09	90.76 ±0.15	86.82 ±0.15	78.83
BGRL	79.43 ±1.24	64.46 ±1.14	83.91 ±0.26	92.63 ±0.21	92.76 ±0.21	88.98 ±0.16	83.70
FUG	82.79 ±0.72	70.52 ±0.15	84.42 ±0.23	92.41 ±0.02	92.31 ±0.16	87.30 ±0.16	84.96
TIG	84.39 ±1.08	73.08 ±0.52	86.45 ±0.13	93.42 ±0.05	93.27 ±0.22	89.51 ±0.07	86.69
GFT	72.99 ±0.94	64.79 ±0.99	83.59 ±0.12	93.06 ±0.27	90.25 ±0.18	84.18 ±0.19	81.48

Appendix

Table 7. Standard Q1 classification accuracy (%, mean ±SD ) over three 20%/80% labeled/test splits. Avg. denotes the six-target mean. Bold and underlined values mark the best and runner-up.

	Node classification (Acc.)				Link prediction (AUC)				Graph classification (Acc.)
Method	Cora	CiteS.	PubM.	Avg.	Phys.	Wiki.	DBLP	Avg.	IMDB	COLL.	PROT.	Avg.
DGI	52.14 ±3.25	46.08 ±3.40	56.76 ±4.35	51.66	98.00 ±0.03	98.81 ±0.04	91.15 ±0.29	95.99	69.62 ±2.01	75.82 ±1.04	71.20 ±2.41	72.21
GRACE	47.83 ±3.33	33.69 ±3.08	51.26 ±4.37	44.26	97.67 ±0.03	98.72 ±0.04	90.36 ±0.41	95.58	68.87 ±2.26	74.72 ±1.17	73.06 ±1.16	72.22
BGRL	56.29 ±3.56	36.06 ±3.09	55.54 ±4.58	49.30	98.46 ±0.03	98.95 ±0.05	91.77 ±0.18	96.39	69.11 ±2.35	75.33 ±1.47	72.49 ±1.76	72.31
FUG	57.19 ±5.58	46.79 ±4.51	64.03 ±5.30	56.00	98.15 ±0.06	98.71 ±0.06	92.46 ±0.19	96.44	68.45 ±2.41	75.53 ±1.06	71.78 ±2.65	71.92
TIG	65.87 ±4.01	60.74 ±3.28	62.03 ±4.75	62.88	98.56 ±0.05	90.62 ±0.28	93.82 ±0.15	94.33	68.94 ±2.38	76.34 ±1.60	72.47 ±2.52	72.58

Appendix

Table 8. Q3 transfer from ogbn-arxiv under the same protocol as Table 3 . Entries are accuracy or AUC (%, mean ±SD ); each Avg. weights three targets equally. Bold and underlined values mark the best and runner-up.

Figure 7. Standard Q1 hyperparameter sensitivity (

k_{\mathrm{train}}=16

): (a) target context size

k

and (b) context weight share

\tau

. Cells report accuracy (%) over three splits; colors show shifts (

\Delta

, pp) from the outlined defaults (

k=12

\tau=0.8

). Two annotated heatmaps showing per-dataset accuracy and average accuracy across target context sizes and context weight shares.

Graph Foundation Models

Graph Foundation Models are pretrained graph-learning systems designed to transfer across datasets, tasks, and domains with different structures and features.

See all 101 papers

Oct 10, 2026

More Domains, No Gains? FLO: Fast-Slow Reciprocal Learning for Multi-Domain Graph Pre-training

Inspired by the great success of foundation models in the development of artificial general intelligence, Graph Foundation Models (GFMs) have recently attracted significant attention and have the potential to advance graph intelligence. Exhibiting such intelligence requires effectively learning knowledge from large-scale, multi-domain graph data during pre-training. Increasing domain diversity during pre-training is expected to yield richer and more generalizable knowledge. Most existing methods pursue this goal by designing domain alignment strategies that help a single graph model learn from multiple domains. However, we observe that existing methods struggle to benefit from increasing domain diversity. Through empirical and theoretical analysis, we reveal two key limitations. Firstly, pre-training in most GFMs is confined to a single message-passing paradigm. Since propagation flattens relations among neighboring nodes, the graph model can hardly perceive the knowledge beyond propagation, namely the attribute-predictable component of these relations. Secondly, domain alignment strategies often rely on randomly initialized parameters that are jointly updated with the graph model using immediate loss feedback, leaving early knowledge acquisition vulnerable to cross-domain interference. To address these issues, we propose FLO, a Fast--sLow reciprOcal approach for multi-domain graph pre-training. Specifically, a propagation-free attribute branch predicts the graph model's relations without message passing and returns these predictions to refine the model. We further stabilize knowledge acquisition by pairing a fast process that adapts to the current domain with a slow process that integrates knowledge across domains. On 12 datasets, FLO benefits from increasing domain diversity and achieves the best average performance in cross-domain and cross-task transfer.

Yupeng Tong, Lianze Shan, Jitao Zhao +2

School of Computer Science and Technology, Tianjin University, Tianjin, China

Graph Foundation Models Cross-Domain Transfer Learning

Ask PDF

Mar 29, 2026

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

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.

Xuanze Chen, Jiajun Zhou, Yadong Li +2

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

Graph Representation Learning Graph Foundation Models Few-Shot Domain Adaptation Graph SSL Prompt Learning

Ask PDF

Jun 16, 2026

Handling Feature Heterogeneity with Learnable Graph Patches

In recent years, the rapid development of foundation models and graph pre-training technologies has spurred increasing interest in constructing a universal pre-trained graph model or Graph Foundation Model (GFM). However, a significant challenge is that existing models are unable to address feature heterogeneity in graph data without textual information, which hinders the transferability of graph models across different datasets. To bridge this gap, we propose the concept of learnable graph patches, which we regard as the smallest semantic units of any graph data. We decompose the graph into learnable graph patches by unfolding the node features and constructing corresponding patch structures separately. We then design a framework that mines transferable information from graph data across domains. Specifically, after extracting graph patches, we propose a patch encoder to extract knowledge from each unit and a patch aggregator to learn how the units are combined into a whole. Due to its domain-agnostic nature, the model can be applied to downstream data across different domains. Furthermore, we analyze the connection between our method and existing graph models, as well as the transferability of the node embeddings it generates. Empirically, our method not only achieves the capability to use multi-domain graphs for pre-training, but also shows enhanced performance across various downstream datasets and tasks. Moreover, we observe consistent improvement in downstream performance as the volume of pre-training data increases.

Yifei Sun, Yang Yang, Xiao Feng +4

Zhejiang University Hangzhou, China · Huazhong University of Science and Technology Wuhan, China · Finvolution Group Shanghai, China

Graph Representation Learning Cross-Domain Transfer Learning Graph Foundation Models

Ask PDF

Abstract

Graph Foundation Models

Explore similar work

More Domains, No Gains? FLO: Fast-Slow Reciprocal Learning for Multi-Domain Graph Pre-training

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

Handling Feature Heterogeneity with Learnable Graph Patches