DualGSL: Prior-Guided Graph Structure Learning with Node-Adaptive Prediction Fusion

Yuxia Sun; Yutszyuk Wong; Huihong Chen; Longhao Lv; Chang-Dong Wang; Li Zeng

cs.LGOct 23, 2026

DualGSL: Prior-Guided Graph Structure Learning with Node-Adaptive Prediction Fusion

Authors: Yuxia Sun, Yutszyuk Wong, Huihong Chen, Longhao Lv, Chang-Dong Wang, Li Zeng

Organizations: the College of Information Science and Technology, Jinan University, Guangzhou 510632, China · Jinan University, Guangzhou 510632, China · Sun Yat-sen University · Changsha University of Science and Technology

Abstract

Graph structure learning (GSL) adapts propagation graphs to prediction tasks, but the relative predictive value of feature-derived priors and learned relations can vary across nodes. We propose DualGSL, a prior-guided framework for transductive node classification that learns node-specific weights for combining predictions from separate prior and task-adaptive propagation paths. The prior branch augments a feature graph with virtual class nodes to incorporate training-label information, while the structural branch combines learned affinity graphs with the feature-graph prior. The two branches share encoder parameters while preserving distinct representations and classification logits. A node-adaptive gate fuses the branch logits based on their projected representations, element-wise absolute difference, and prediction confidence. A joint objective integrates fused and branch-specific classification losses with prior-branch class-anchor alignment, coordinating graph construction, representation learning, and prediction fusion. Experiments against seven baselines on the AMG18 malware attribution dataset and the widely used 20News text classification and USPS handwritten-digit recognition benchmarks show that DualGSL achieves the highest cross-dataset average accuracy and Macro-F1 among the evaluated methods. Ablation studies on AMG18 further demonstrate the benefit of combining prior and learned propagation through logit fusion.

Figures & tables

Fig. 1: Overview of the DualGSL architecture. The label-augmented prior graph $\widetilde{\mathbf{A}}_{p}$ and the learned structural graph $\mathbf{A}_{c}$ define separate propagation paths with shared encoder parameters and branch-specific classification heads. Projected representations, their element-wise absolute difference, and prediction confidence determine the node-wise gate $g_{i}$ . The global scale $\sigma(\beta)$ gives the structural mixing coefficient $\gamma_{i}=\sigma(\beta)g_{i}$ for combining the two branch logits. During training, source-node labels construct virtual-class connections, while prediction-node labels supervise the classification and class-anchor alignment objectives.

Dataset	Nodes	Features	Classes	Train	Val	Test
AMG18	6,360	128	18	4,573	509	1,278
20News	18,846	768	20	10,182	1,132	7,532
USPS	9,298	256	10	6,561	730	2,007

TABLE I: Dataset statistics after validation splitting.

Dataset	$k$	Source ratio	$\lambda_{\mathrm{center}}$	$\lambda_{\mathrm{view}}$	Seeds
AMG18	10	0.5	0.05	0.40	1–5
20News	5	0.5	0.05	0.05	1–5
USPS	3	0.5	0.05	–	1–5

TABLE II: Dataset-specific hyperparameters used for the main results. Each model is trained using Algorithm 1 with five random seeds. A dash indicates an inactive term.

	AMG18		20News		USPS		Cross-dataset average
Method	Acc.	Macro-F1	Acc.	Macro-F1	Acc.	Macro-F1	Avg. Acc.	Avg. Macro-F1
MLP	$\underline{88.65}\pm 0.31$	$\underline{86.85}\pm 0.72$	$71.94\pm 0.13$	$70.66\pm 0.08$	$93.16\pm 0.10$	$92.61\pm 0.14$	$84.58$	$83.37$
GCN [ 1 ]	$88.29\pm 0.14$	$86.29\pm 0.11$	$\mathbf{73.59}\pm 0.13$	$\mathbf{72.30}\pm 0.12$	$93.54\pm 0.21$	$93.13\pm 0.22$	$85.14$	$83.91$
GSR [ 7 ]	$\underline{88.65}\pm 0.21$	$86.68\pm 0.31$	$72.48\pm 0.27$	$71.32\pm 0.29$	$94.24\pm 0.19$	$93.77\pm 0.21$	$85.12$	$\underline{83.92}$
IDGL [ 5 ]	$82.46\pm 0.59$	$77.98\pm 1.07$	$70.03\pm 0.27$	$68.56\pm 0.38$	$91.66\pm 0.30$	$90.92\pm 0.36$	$81.38$	$79.15$
ELU-GCN [ 11 ]	$86.90\pm 0.27$	$84.51\pm 0.26$	$\underline{72.92}\pm 0.05$	$71.44\pm 0.18$	$93.38\pm 0.24$	$92.99\pm 0.26$	$84.40$	$82.98$

TABLE III: Classification performance (%). Per-dataset results are reported as mean

\pm

standard deviation over five random seeds. The last two columns report the unweighted averages of the three per-dataset means. Bold and underlined values indicate the best and second-best means in each column, respectively.

Variant	Acc. (%)	$\Delta$ Acc. (pp)	Macro-F1 (%)	$\Delta$ Macro-F1 (pp)
Full model	88.76	—	87.28	—
w/o structural branch	87.87	$-0.89$	85.93	$-1.35$
w/o prior branch	87.37	$-1.39$	84.88	$-2.40$
w/o graph learners	88.09	$-0.67$	86.04	$-1.24$
w/o virtual class nodes	88.56	$-0.20$	86.88	$-0.40$
w/o center alignment	88.59	$-0.17$	86.98	$-0.30$

TABLE IV: Ablation study on AMG18. Results are averaged over five random seeds.

Prediction	Acc. (%)	Macro-F1 (%)
Prior branch	87.72	86.52
Structural branch	87.32	85.95
Fused prediction	88.73	88.16

TABLE V: Prediction fusion on AMG18 at the same trained dual-view checkpoint. Bold values indicate the best results.

Graph / subset	# Edges	SC edges (%)	SC weight (%)
Prior graph $\mathbf{A}_{p}$	89,756	93.21	93.09
Structural graph $\mathbf{A}_{c}$	283,647	53.44	68.68
Inherited edges in $\mathbf{A}_{c}$	89,756	93.21	95.68
New edges in $\mathbf{A}_{c}$	193,891	35.04	41.24

TABLE VI: Edge counts and class consistency on AMG18. # Edges denotes the number of positive-weight, non-self directed edges between real nodes. SC denotes same-class. For each row, SC edges reports the percentage of edges joining nodes with the same class label, and SC weight reports the percentage of total edge weight carried by these same-class edges. Weights are taken from

\mathbf{A}_{p}

for the first row and from

\mathbf{A}_{c}

for the remaining rows.

Graph condition	Removed weight (%)	Acc. (%)	Macro-F1 (%)
Full $\mathbf{A}_{c}$	0.00	88.73	88.16
New edges removed	49.59	88.97	88.01
Random edges removed	69.14	88.58	87.65

TABLE VII: Edge-removal analysis on AMG18 at a fixed checkpoint. Removed weight is the percentage of the full

\mathbf{A}_{c}

’s non-self edge weight removed, computed before deletion and renormalization. Bold values indicate the best prediction results.

Fig. 2: Visualization of input features, branch representations, and structural mixing coefficients on AMG18. Panels (a)–(c) show two-dimensional projections of input features, prior-branch representations $\mathbf{Z}_{p}$ , and structural-branch representations $\mathbf{Z}_{c}$ , respectively, with nodes colored consistently by APT group. Panel (d) uses the same coordinates as panel (b), with node colors indicating the structural mixing coefficients $\gamma_{i}$ .

Parameter	Value	Acc. (%)	Macro-F1 (%)
$k$	3	$88.65\pm 0.73$	$86.79\pm 0.71$
$k$	5	$\mathbf{89.22}\pm 0.40$	$\mathbf{87.63}\pm 0.62$
$k$	8	$88.53\pm 0.62$	$86.89\pm 0.76$
$k$	$10^{\dagger}$	$88.76\pm 0.23$	$87.28\pm 0.64$
$k$	15	$88.95\pm 0.77$	$87.21\pm 0.99$
$\lambda_{\mathrm{view}}$	0.00	$87.93\pm 0.31$	$86.32\pm 0.70$

TABLE VIII: Hyperparameter sensitivity on AMG18. Results are mean

\pm

standard deviation over five random seeds. Bold means indicate the best result for each metric within each sweep. † marks the main-experiment settings; the corresponding rows report the same five AMG18 runs as Table III .

Graph Structure Learning

Graph structure learning infers or adapts network connections from data for statistical discovery and downstream prediction.

See all 139 papers

PreGS: A Parameter-Transfer-Based Multi-Expert Graph Neural Network for Node Classification
Sep 22, 2026Zhicong Cai, Yinglong Zhang, Xiaoying Hong +2Graph Neural Networks Mixture-of-Experts Models
College of Physics and Information Engineering, Minnan Normal University, Zhangzhou 363000, China
Structure-Feature Aligned Graph Learning via Alternating Constrained Optimization
Jul 13, 2026Chengcheng Yan, Qingsong WangGraph Neural Networks Graph Representation Learning
College of Artificial Intelligence, Shaoxing Institute of Technology, Shaoxing, 312000, China. · School of Mathematics and Computational Science, Xiangtan University, Xiangtan, 411105, China.
Learning Posterior Predictive Distributions for Node Classification from Synthetic Graph Priors
Apr 21, 2026Jeongwhan Choi, Jongwoo Kim, Woosung Kang +1Synthetic Data Pretraining OOD Generalization
KAIST Daejeon, Republic of Korea

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

PreGS: A Parameter-Transfer-Based Multi-Expert Graph Neural Network for Node Classification

Zhicong Cai, Yinglong Zhang, Xiaoying Hong, Xuewen Xia+1

College of Physics and Information Engineering, Minnan Normal University, Zhangzhou 363000, China

Graph Neural Networks Mixture-of-Experts Models Node Classification

Open paper AskPDF arXiv

Graph neural networks have achieved strong performance in node classification by aggregating information from graph neighborhoods. However, a single aggregation mechanism may be insufficient to capture diverse structural patterns across graph datasets. Moreover, independently training multiple structural branches can introduce substantial overhead without necessarily producing stable node representations. To address these issues, this paper proposes PreGS, a parameter-transfer-based multi-expert graph neural network framework. PreGS first pretrains a multi-head graph attention network (GAT) and transfers the linear transformation weights of its first-layer attention heads to multiple GraphSAGE experts. The transferred experts are frozen and used as complementary structural branches. The fused raw node features, GAT head representations, and GraphSAGE expert representations are fed into a multilayer perceptron (MLP), whose output is further fused with the pretrained GAT logits. Based on PreGS, we further develop PreGSv2, which introduces source-level weighting and a structural gating mechanism for adaptive multi-source feature integration. Experiments on eight public graph datasets show that PreGS and PreGSv2 achieve competitive performance against representative graph neural network baselines. Ablation, parameter-transfer, sensitivity, aggregator, visualization, and training-time analyses further validate the effectiveness and stability of the proposed framework. The code and datasets are available at https://github.com/LH-Czc/PreGS.

Abstract

Graph Structure Learning

Explore similar work

PreGS: A Parameter-Transfer-Based Multi-Expert Graph Neural Network for Node Classification

Structure-Feature Aligned Graph Learning via Alternating Constrained Optimization

Learning Posterior Predictive Distributions for Node Classification from Synthetic Graph Priors