Data-Efficient Agentic Graph Domain Adaptation via Reliability-Aware Prototype Learning

Yingxu Wang; Kunyu Zhang; Siyang Gao

cs.LGSep 12, 2026

Data-Efficient Agentic Graph Domain Adaptation via Reliability-Aware Prototype Learning

Authors: Yingxu Wang, Kunyu Zhang, Siyang Gao

Organizations: Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE · Zhengzhou University, Zhengzhou, China · City University of Hong Kong, Hong Kong SAR, China

Abstract

Agentic learning systems are often required to adapt after deployment by observing new data and reusing prior knowledge under limited supervision or feedback. For graph-structured prediction, Graph Domain Adaptation (GDA) naturally instantiates this setting by transferring knowledge from labeled source graphs to unlabeled target graphs under distribution shifts. However, most GDA methods assume sufficient labeled source graphs, which becomes restrictive in data-efficient agentic settings where only limited source evidence can be retained. Under such constraints, source semantics become unreliable, leading to unstable source anchoring, uncertain target association, and fragile targetmarginal calibration. To address these challenges, we propose DEAG, a reliability-aware prototype learning framework for data-efficient agentic GDA. DEAG estimates class reliability from retained source support and embedding compactness, and constructs stable reusable source anchors by blending empirical prototypes with classifier directions. Guided by these anchors, DEAG performs prototype-aware soft target association and aligns confidence-weighted target centers with source semantics. A source-prior regularizer further sharpens target predictions while keeping the target marginal consistent with retained source evidence. Experiments on graph benchmarks with diverse domain shifts show that DEAG improves average adaptation performance over competitive GDA baselines under the same source-data budget.

Figures & tables

Figure 1: Overview of the proposed DEAG . DEAG estimates source-class reliability using support size and embedding compactness, constructs reusable source anchors by blending source prototypes with classifier directions, softly associates target graphs with these anchors, and jointly optimizes prototype alignment and source-prior target-marginal calibration.

Methods	Node Shift			Edge Shift			Feature Shift
Methods	M0 $\rightarrow$ M1	M0 $\rightarrow$ M2	M0 $\rightarrow$ M3	M0 $\rightarrow$ M1	M0 $\rightarrow$ M2	M0 $\rightarrow$ M3	P $\rightarrow$ D	D $\rightarrow$ P	C $\rightarrow$ CM	CM $\rightarrow$ C	B $\rightarrow$ BM	BM $\rightarrow$ B
GCN	53.2 $\pm$ 4.1	49.5 $\pm$ 3.6	50.1 $\pm$ 4.5	61.5 $\pm$ 5.4	53.2 $\pm$ 4.8	47.3 $\pm$ 3.2	52.3 $\pm$ 4.1	48.7 $\pm$ 5.2	41.2 $\pm$ 3.8	45.6 $\pm$ 2.9	44.3 $\pm$ 6.2	43.8 $\pm$ 4.7
GIN	51.4 $\pm$ 5.2	47.8 $\pm$ 4.4	49.6 $\pm$ 3.8	59.8 $\pm$ 6.1	51.6 $\pm$ 5.2	46.5 $\pm$ 4.1	50.7 $\pm$ 5.3	47.2 $\pm$ 4.6	40.8 $\pm$ 4.1	44.9 $\pm$ 3.2	43.6 $\pm$ 5.8	42.1 $\pm$ 4.9
GMT	55.1 $\pm$ 3.8	51.2 $\pm$ 3.2	51.5 $\pm$ 2.9	62.4 $\pm$ 4.5	54.5 $\pm$ 3.9	48.2 $\pm$ 2.8	54.2 $\pm$ 3.7	49.6 $\pm$ 3.4	42.4 $\pm$ 3.6	46.8 $\pm$ 2.8	45.4 $\pm$ 5.1	44.7 $\pm$ 4.2
CIN	54.3 $\pm$ 4.5	50.6 $\pm$ 3.7	50.9 $\pm$ 3.4	61.9 $\pm$ 5.0	53.8 $\pm$ 4.2	47.8 $\pm$ 3.1	53.6 $\pm$ 4.2	48.9 $\pm$ 4.1	41.7 $\pm$ 3.9	45.2 $\pm$ 3.1	44.8 $\pm$ 5.4	43.9 $\pm$ 4.5
PathNN	52.8 $\pm$ 4.8	48.9 $\pm$ 4.1	50.4 $\pm$ 3.6	60.5 $\pm$ 5.6	52.4 $\pm$ 4.9	46.9 $\pm$ 3.7	51.8 $\pm$ 4.6	47.8 $\pm$ 4.3	41.1 $\pm$ 4.2	44.7 $\pm$ 3.4	43.9 $\pm$ 5.6	43.2 $\pm$ 4.8

Table 1: Graph classification results (in %) under node and edge density domain shifts on the Mutagenicity dataset, and under feature shifts (source

\rightarrow

target). P, D, C, CM, B, and BM denote PROTEINS, DD, COX2, COX2_MD, BZR, and BZR_MD, respectively. Bold indicates the best performance.

Methods	Node Avg.	Edge Avg.	Feature Avg.	Avg. Rank	$p$ -value	W/T/L
GCN	46.82	47.23	45.98	11.52	$1.84{\times}10^{-23}$	54/0/0
GIN	46.25	46.87	44.88	12.34	$1.23{\times}10^{-23}$	54/0/0
GMT	48.03	48.46	47.18	7.78	$4.25{\times}10^{-22}$	54/0/0
CIN	47.50	47.89	46.35	9.91	$8.96{\times}10^{-23}$	54/0/0
PathNN	47.08	47.52	45.42	10.13	$4.90{\times}10^{-23}$	54/0/0
DEAL	57.10	56.87	49.67	3.75	$3.38{\times}10^{-15}$	52/1/1

Table 2: Aggregate comparison across all reported tasks. Bold results indicate the best performance.

Panel of Figure 2 — Figure 2: Information yield under source-data budgets and sensitivity analysis on Mutagenicity.

Method	M0 $\rightarrow$ M1	M0 $\rightarrow$ M2	M0 $\rightarrow$ M3	M1 $\rightarrow$ M0	M1 $\rightarrow$ M2	M1 $\rightarrow$ M3	M2 $\rightarrow$ M0	M2 $\rightarrow$ M1	M2 $\rightarrow$ M3	M3 $\rightarrow$ M0	M3 $\rightarrow$ M1	M3 $\rightarrow$ M2
DEAG w/o $\mathcal{L}_{pa}$	70.8	63.4	57.1	65.4	67.5	59.8	62.3	73.1	61.2	58.6	60.1	67.8
DEAG w/o $\mathcal{L}_{ptr}$	73.4	66.1	60.3	67.3	70.2	62.5	65.0	75.8	64.1	62.1	63.4	70.6
DEAG w/o $H_{\text{cond}}$	74.9	67.5	61.6	69.8	71.8	64.1	66.9	77.4	65.7	63.4	65.0	72.3
DEAG w/o $D_{\text{prior}}$	75.3	68.2	61.2	70.7	72.3	64.6	67.8	77.8	66.2	63.1	65.5	72.7
DEAG	76.9	69.7	63.4	71.8	73.7	65.9	69.1	79.2	67.5	65.0	66.8	74.0

Table 3: The results of ablation studies on the Mutagenicity dataset. Bold results indicate the best performance.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Datasets	Graphs	Avg. Nodes	Avg. Edges	Classes
NCI1	4,110	29.87	32.30	2
Mutagenicity	4,337	30.32	30.77	2
DD	1,178	284.32	715.66	2
PROTEINS	1,113	39.1	72.8	2
COX2	467	41.22	43.45	2
COX2_MD	303	26.28	335.12	2

Appendix

Table 4: Statistics of the experimental datasets.

Method	M0 $\rightarrow$ M1	M0 $\rightarrow$ M2	M0 $\rightarrow$ M3	M1 $\rightarrow$ M0	M1 $\rightarrow$ M2	M1 $\rightarrow$ M3	M2 $\rightarrow$ M0	M2 $\rightarrow$ M1	M2 $\rightarrow$ M3	M3 $\rightarrow$ M0	M3 $\rightarrow$ M1	M3 $\rightarrow$ M2
GCN	53.2 $\pm$ 4.1	49.5 $\pm$ 3.6	50.1 $\pm$ 4.5	46.5 $\pm$ 2.4	58.2 $\pm$ 2.1	45.7 $\pm$ 1.6	46.1 $\pm$ 1.8	65.3 $\pm$ 2.5	45.1 $\pm$ 1.9	48.6 $\pm$ 2.8	41.2 $\pm$ 6.2	43.5 $\pm$ 5.1
GIN	51.4 $\pm$ 5.2	47.8 $\pm$ 4.4	49.6 $\pm$ 3.8	45.1 $\pm$ 2.9	56.8 $\pm$ 2.5	44.3 $\pm$ 1.9	44.8 $\pm$ 2.1	63.9 $\pm$ 3.1	43.7 $\pm$ 2.2	47.1 $\pm$ 3.1	39.5 $\pm$ 5.8	41.8 $\pm$ 6.5
GMT	55.1 $\pm$ 3.8	51.2 $\pm$ 3.2	51.5 $\pm$ 2.9	47.8 $\pm$ 2.1	59.4 $\pm$ 1.8	46.8 $\pm$ 1.4	47.3 $\pm$ 1.5	66.8 $\pm$ 2.2	46.3 $\pm$ 1.6	49.5 $\pm$ 2.4	42.6 $\pm$ 5.5	44.7 $\pm$ 4.8
CIN	54.3 $\pm$ 4.5	50.6 $\pm$ 3.7	50.9 $\pm$ 3.4	47.2 $\pm$ 2.3	58.9 $\pm$ 2.0	46.2 $\pm$ 1.5	46.7 $\pm$ 1.7	66.1 $\pm$ 2.4	45.8 $\pm$ 1.8	48.9 $\pm$ 2.9	41.9 $\pm$ 6.0	43.9 $\pm$ 5.6
PathNN	52.8 $\pm$ 4.8	48.9 $\pm$ 4.1	50.4 $\pm$ 3.6	45.8 $\pm$ 2.7	57.5 $\pm$ 2.4	45.1 $\pm$ 1.8	45.4 $\pm$ 2.0	64.5 $\pm$ 2.8	44.6 $\pm$ 2.1	47.8 $\pm$ 3.0	40.7 $\pm$ 5.9	42.6 $\pm$ 6.1
DEAL	72.1 $\pm$ 2.1	64.8 $\pm$ 2.4	56.1 $\pm$ 1.2	67.8 $\pm$ 2.2	68.0 $\pm$ 3.1	58.5 $\pm$ 2.8	69.1 $\pm$ 2.9	72.4 $\pm$ 1.3	61.2 $\pm$ 2.2	59.4 $\pm$ 3.9	52.5 $\pm$ 2.3	65.4 $\pm$ 4.1

Appendix

Table 5: The classification results (in %) on the Mutagenicity dataset under node density domain shift (source → target). M0, M1, M2, and M3 denote the sub-datasets partitioned with node density. Bold results indicate the best performance.

Method	M0 $\rightarrow$ M1	M0 $\rightarrow$ M2	M0 $\rightarrow$ M3	M1 $\rightarrow$ M0	M1 $\rightarrow$ M2	M1 $\rightarrow$ M3	M2 $\rightarrow$ M0	M2 $\rightarrow$ M1	M2 $\rightarrow$ M3	M3 $\rightarrow$ M0	M3 $\rightarrow$ M1	M3 $\rightarrow$ M2
GCN	61.5 $\pm$ 5.4	53.2 $\pm$ 4.8	47.3 $\pm$ 3.2	50.1 $\pm$ 2.5	55.4 $\pm$ 2.8	45.3 $\pm$ 1.9	50.5 $\pm$ 2.2	64.2 $\pm$ 3.5	45.0 $\pm$ 1.7	47.9 $\pm$ 2.5	33.8 $\pm$ 2.7	42.1 $\pm$ 4.6
GIN	59.8 $\pm$ 6.1	51.6 $\pm$ 5.2	46.5 $\pm$ 4.1	48.7 $\pm$ 3.1	54.1 $\pm$ 3.4	44.2 $\pm$ 2.1	49.3 $\pm$ 2.5	62.8 $\pm$ 4.1	43.8 $\pm$ 2.0	46.2 $\pm$ 3.3	32.6 $\pm$ 3.5	41.5 $\pm$ 5.2
GMT	62.4 $\pm$ 4.5	54.5 $\pm$ 3.9	48.2 $\pm$ 2.8	51.5 $\pm$ 2.2	56.6 $\pm$ 2.4	46.4 $\pm$ 1.5	51.8 $\pm$ 1.8	65.5 $\pm$ 3.0	46.1 $\pm$ 1.4	48.7 $\pm$ 2.1	34.5 $\pm$ 2.9	43.8 $\pm$ 3.7
CIN	61.9 $\pm$ 5.0	53.8 $\pm$ 4.2	47.8 $\pm$ 3.1	50.8 $\pm$ 2.4	55.9 $\pm$ 2.7	45.8 $\pm$ 1.7	51.1 $\pm$ 2.1	64.8 $\pm$ 3.3	45.5 $\pm$ 1.6	48.1 $\pm$ 2.4	34.1 $\pm$ 3.1	42.7 $\pm$ 4.1
PathNN	60.5 $\pm$ 5.6	52.4 $\pm$ 4.9	46.9 $\pm$ 3.7	49.3 $\pm$ 2.8	54.8 $\pm$ 3.1	44.9 $\pm$ 2.0	49.8 $\pm$ 2.4	63.5 $\pm$ 3.8	44.5 $\pm$ 1.9	47.3 $\pm$ 2.8	33.2 $\pm$ 3.2	41.9 $\pm$ 4.8
DEAL	72.8 $\pm$ 0.7	62.5 $\pm$ 2.1	54.2 $\pm$ 1.4	67.5 $\pm$ 2.8	64.2 $\pm$ 4.9	58.6 $\pm$ 2.1	68.9 $\pm$ 1.8	66.4 $\pm$ 3.5	60.1 $\pm$ 4.1	57.8 $\pm$ 1.7	55.7 $\pm$ 6.2	66.1 $\pm$ 2.5

Appendix

Table 6: The classification results (in %) on the Mutagenicity dataset under edge density domain shift (source → target). M0, M1, M2, and M3 denote the sub-datasets partitioned with edge density. Bold results indicate the best performance.

Method	N0 $\rightarrow$ N1	N0 $\rightarrow$ N2	N0 $\rightarrow$ N3	N1 $\rightarrow$ N0	N1 $\rightarrow$ N2	N1 $\rightarrow$ N3	N2 $\rightarrow$ N0	N2 $\rightarrow$ N1	N2 $\rightarrow$ N3	N3 $\rightarrow$ N0	N3 $\rightarrow$ N1	N3 $\rightarrow$ N2
GCN	49.5 $\pm$ 2.1	42.2 $\pm$ 1.8	32.1 $\pm$ 2.4	62.1 $\pm$ 15.5	41.2 $\pm$ 1.8	38.5 $\pm$ 11.2	25.5 $\pm$ 1.4	47.5 $\pm$ 1.5	61.2 $\pm$ 1.1	26.5 $\pm$ 1.5	48.3 $\pm$ 1.7	56.1 $\pm$ 2.2
GIN	50.1 $\pm$ 1.9	42.4 $\pm$ 2.0	31.8 $\pm$ 1.9	62.6 $\pm$ 16.1	41.6 $\pm$ 2.1	38.9 $\pm$ 11.8	25.8 $\pm$ 1.6	47.9 $\pm$ 1.7	61.6 $\pm$ 1.3	26.8 $\pm$ 1.3	48.8 $\pm$ 2.1	55.8 $\pm$ 2.4
GMT	51.2 $\pm$ 2.4	43.1 $\pm$ 2.5	33.0 $\pm$ 2.8	63.8 $\pm$ 14.8	42.5 $\pm$ 1.5	39.8 $\pm$ 10.5	26.4 $\pm$ 1.2	48.8 $\pm$ 1.3	62.5 $\pm$ 0.9	27.0 $\pm$ 1.6	49.1 $\pm$ 1.8	56.5 $\pm$ 1.9
CIN	50.6 $\pm$ 2.1	42.8 $\pm$ 2.3	32.5 $\pm$ 2.2	63.2 $\pm$ 15.3	41.9 $\pm$ 1.7	39.2 $\pm$ 11.0	26.1 $\pm$ 1.3	48.2 $\pm$ 1.4	61.9 $\pm$ 1.0	26.9 $\pm$ 1.8	49.0 $\pm$ 2.0	56.2 $\pm$ 2.1
PathNN	51.5 $\pm$ 1.8	43.5 $\pm$ 1.9	33.2 $\pm$ 2.0	63.5 $\pm$ 15.0	42.2 $\pm$ 1.9	39.5 $\pm$ 10.8	26.3 $\pm$ 1.5	48.5 $\pm$ 1.6	62.2 $\pm$ 1.2	27.1 $\pm$ 1.2	49.4 $\pm$ 1.5	56.8 $\pm$ 1.7
DEAL	53.2 $\pm$ 1.8	46.1 $\pm$ 3.6	38.8 $\pm$ 6.2	71.1 $\pm$ 1.5	54.9 $\pm$ 3.5	48.4 $\pm$ 2.4	40.2 $\pm$ 12.5	50.3 $\pm$ 2.8	66.5 $\pm$ 0.9	27.2 $\pm$ 1.1	49.3 $\pm$ 1.2	57.2 $\pm$ 1.8

Appendix

Table 7: The classification results (in %) on the NCI1 dataset under node density domain shift (source → target). N0, N1, N2, and N3 denote the sub-datasets partitioned with node density. Bold results indicate the best performance.

Method	N0 $\rightarrow$ N1	N0 $\rightarrow$ N2	N0 $\rightarrow$ N3	N1 $\rightarrow$ N0	N1 $\rightarrow$ N2	N1 $\rightarrow$ N3	N2 $\rightarrow$ N0	N2 $\rightarrow$ N1	N2 $\rightarrow$ N3	N3 $\rightarrow$ N0	N3 $\rightarrow$ N1	N3 $\rightarrow$ N2
GCN	47.8 $\pm$ 1.9	45.3 $\pm$ 2.1	30.8 $\pm$ 1.6	43.1 $\pm$ 15.2	48.2 $\pm$ 2.4	51.5 $\pm$ 15.2	34.2 $\pm$ 16.5	49.5 $\pm$ 1.5	58.7 $\pm$ 8.5	26.8 $\pm$ 1.4	49.5 $\pm$ 1.8	51.9 $\pm$ 2.0
GIN	48.2 $\pm$ 1.7	46.1 $\pm$ 1.5	31.2 $\pm$ 2.1	43.5 $\pm$ 16.1	48.7 $\pm$ 2.7	52.1 $\pm$ 16.1	34.8 $\pm$ 17.1	50.1 $\pm$ 1.7	59.3 $\pm$ 9.1	27.1 $\pm$ 1.8	50.2 $\pm$ 2.2	52.4 $\pm$ 1.6
GMT	49.9 $\pm$ 2.2	46.5 $\pm$ 2.4	31.5 $\pm$ 1.9	44.8 $\pm$ 14.8	49.5 $\pm$ 2.1	53.0 $\pm$ 14.8	35.6 $\pm$ 15.8	51.0 $\pm$ 1.3	60.5 $\pm$ 7.8	27.2 $\pm$ 1.5	50.8 $\pm$ 1.7	52.8 $\pm$ 2.1
CIN	48.5 $\pm$ 1.8	46.2 $\pm$ 2.0	31.3 $\pm$ 1.7	44.2 $\pm$ 15.3	48.9 $\pm$ 2.3	52.4 $\pm$ 15.3	35.1 $\pm$ 16.2	50.4 $\pm$ 1.4	59.9 $\pm$ 8.2	27.0 $\pm$ 1.7	50.5 $\pm$ 1.6	52.6 $\pm$ 1.9
PathNN	49.1 $\pm$ 1.5	46.8 $\pm$ 1.8	31.7 $\pm$ 1.4	44.6 $\pm$ 15.0	49.2 $\pm$ 2.5	52.8 $\pm$ 15.0	35.4 $\pm$ 16.0	50.7 $\pm$ 1.6	60.2 $\pm$ 8.0	27.3 $\pm$ 1.3	50.6 $\pm$ 1.7	53.0 $\pm$ 1.5
DEAL	52.1 $\pm$ 2.7	49.2 $\pm$ 1.8	36.7 $\pm$ 6.4	70.8 $\pm$ 2.5	53.8 $\pm$ 2.1	54.7 $\pm$ 7.2	42.8 $\pm$ 13.5	54.1 $\pm$ 3.4	64.1 $\pm$ 4.8	27.5 $\pm$ 1.2	51.2 $\pm$ 1.1	53.1 $\pm$ 1.4

Appendix

Table 8: The classification results (in %) on the NCI1 dataset under edge density domain shift (source → target). N0, N1, N2, and N3 denote the sub-datasets partitioned with edge density. Bold results indicate the best performance.

Graph Domain Adaptation

See all 9 papers

Graph Domain Adaptation Does Not End with Representation Learning
Sep 22, 2026Ziqian Liu, Yongxue Xu, Enze Zhang +3Unsupervised Domain Adaptation Graph Representation Learning
Sun Yat-sen University, China · University of Wisconsin–Madison, USA · Jiangsu University, China +2
Cross-Resolution Semantic Learning for Graph Domain Adaptation
Jul 31, 2026Yingxu Wang, Haoze Huang, Zhongkai Zheng +1Unsupervised Domain Adaptation Graph Representation Learning
Mohamed bin Zayed University of Artificial Intelligence · Fuzhou University · Sun Yat-sen University +1
Safe-Subspace Pseudo-Label Refinement for Source-Free Graph Domain Adaptation
May 30, 2026Yingxu Wang, Xinwang Liu, Siyang Gao +1Source-Free Domain Adaptation Pseudo-Labeling
Department of Computer Science and Engineering, Chinese University of Hong Kong · College of Computer, National University of Defense Technology, Changsha, 410073, China · Department of Data Science, City University of Hong Kong +1

Sep 22, 2026J/K move · Enter open · S save

Graph Domain Adaptation Does Not End with Representation Learning

Ziqian Liu, Yongxue Xu, Enze Zhang, Jiaqi Zhang+2

Sun Yat-sen University, China · University of Wisconsin–Madison, USA · Jiangsu University, China+2

Unsupervised Domain Adaptation Graph Representation Learning Domain Adaptation Graph Domain Adaptation Node Classification

Open paper AskPDF arXiv

Graph domain adaptation (GDA) transfers knowledge from a labeled source graph to an unlabeled target graph under shifts in both node attributes and graph structure. Existing methods primarily adapt graph representations through propagation redesign, distribution alignment, or source-to-target transition modeling, but still rely on a single graph-propagating path for target prediction. This leaves open whether an adapted graph representation exhausts the predictive evidence available in the target domain, since the graph-aware expert and graph-free local expert may exhibit different failure modes under topological shifts. To address this limitation, we propose EviGDA, an Evidence-Augmented Graph Domain Adaptation framework that complements graph representation adaptation with a graph-free local expert. The graph-aware expert performs message passing and entropy-aware marginal alignment, while the graph-free local expert learns solely from source node features and labels without graph propagation or target alignment. The two experts are optimized independently and combined only at inference through a task-level constant probability mixture, preserving complementary evidence without joint training, learned routing, or target pseudo-labels. Extensive experiments on ten datasets and 16 transfer tasks show that EviGDA outperforms state-of-the-art baselines. Complementary prediction paths add value beyond graph representation alignment.

Abstract

Graph Domain Adaptation

Explore similar work

Graph Domain Adaptation Does Not End with Representation Learning

Cross-Resolution Semantic Learning for Graph Domain Adaptation

Safe-Subspace Pseudo-Label Refinement for Source-Free Graph Domain Adaptation