BAED: a New Paradigm for Few-shot Graph Learning with Explanation in the Loop

Chao Chen; Xujia Li; Dongsheng Hong; Shanshan Lin; Xiangwen Liao; Chuanyi Liu; Lei Chen

cs.LGMar 2, 2026

BAED: a New Paradigm for Few-shot Graph Learning with Explanation in the Loop

Authors: Chao Chen, Xujia Li, Dongsheng Hong, Shanshan Lin, Xiangwen Liao, Chuanyi Liu, Lei Chen

Organizations: Harbin Institute of Technology (Shenzhen), Guangdong, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China · Fuzhou University, Fujian, China · The Hong Kong University of Science and Technology (Guangzhou), Guangdong, China

Abstract

The challenges of training and inference in few-shot environments persist in the area of graph representation learning. The quality and quantity of labels are often insufficient due to the extensive expert knowledge required to annotate graph data. In this context, Few-Shot Graph Learning (FSGL) approaches have been developed over the years. Through sophisticated neural architectures and customized training pipelines, these approaches enhance model adaptability to new label distributions. However, compromises in \textcolor{black}{the model's} robustness and interpretability can result in overfitting to noise in labeled data and degraded performance. This paper introduces the first explanation-in-the-loop framework for the FSGL problem, called BAED. We novelly employ the belief propagation algorithm to facilitate label augmentation on graphs. Then, leveraging an auxiliary graph neural network and the gradient backpropagation method, our framework effectively extracts explanatory subgraphs surrounding target nodes. The final predictions are based on these informative subgraphs while mitigating the influence of redundant information from neighboring nodes. Extensive experiments on seven benchmark datasets demonstrate superior prediction accuracy, training efficiency, and explanation quality of BAED. As a pioneer, this work highlights the potential of the explanation-based research paradigm in FSGL.

Figures & tables

Figure 1: Explanation in the loop: a new paradigm for FSGL

Notation	Description
$\bm{x}_{i}$	Feature vector of a node $v_{i}$
$c_{i}$	A variable representing one potential class of node $v_{i}$ , $c_{i}\in\mathcal{C}$
$\bm{\hat{c}}_{i}$	Class probability distribution vector of $v_{i}$ predicted by models
$\bm{\phi}_{i}$	Prior probability vector of node $v_{i}$
$\bm{\phi}_{i}(c_{i})$	Prior probability of node $v_{i}$ belonging to one specific class $c_{i}$
$\bm{b}_{i}$	Posterior probability vector of node $v_{i}$

Table 1: Major notations

Figure 2: BAED: the explanation-based FSGL pipeline

Figure 3: Explanatory subgraphs extracted using gradient backpropagation and edge ranking on the PubMed. Colored backgrounds (red, green, blue) indicate ground-truth node labels; gray nodes are unlabeled. Dashed circles denote the explanatory subgraphs for node A, B, and C, with edge numbers indicating importance rankings. Nodes A and B lie within homogeneous regions of same-label nodes, resulting in clean subgraphs without distracting connections. Node C, positioned near both red and blue nodes, yields a mixed but predominantly blue subgraph, leading to a correct prediction.

Datasets	Cora	Citeseer	PubMed	Wiki	DBLP	Wisconsin	CoauthorCS
Random walk	2.5	3.6	2.2	4.2	2.8	4.9	4.4
BAED	94.6	94.3	87.2	92.3	96.8	93.9	95.5

Table 2: Proportion of Subgraphs with Labels

\uparrow

Datasets	# Classes	# Features	# Edges	# Nodes	Avg. Degree
Cora ( Namata et al., 2016 )	7	1,433	5,429	2,708	3.90
Citeseer ( Bojchevski and Günnemann, 2018 )	6	3,703	4,715	3,312	2.85
PubMed ( Dernoncourt and Lee, 2017 )	3	500	88,676	19,717	4.50
Wiki ( Lim et al., 2021 )	17	4,973	17,981	2,405	14.95
DBLP ( Tang et al., 2008 )	3	1,639	105,734	17,716	11.94
Wisconsin ( Pei et al., 2020 )	5	1,703	515	251	4.10

Table 3: Statistics of Datasets

Datasets	Traditional GNNs				Advanced GNNs			FSGL Models			BAED		Improve.
Datasets	SAGE	GAT	GIN	GCN	SGC	DLRGAE	HiD	GPN	tsGCN	DCI	SAGE+IG	SAGE+SM	Improve.
Cora	0.301	0.307	0.269	0.311*	0.301	0.301	0.300	0.251	0.296	0.238	0.515	0.518	66.5%
Citeseer	0.202	0.192	0.215*	0.199	0.189	0.189	0.204	0.199	0.211	0.199	0.769	0.787	266.1%
PubMed	0.400	0.393	0.392	0.396	0.410	0.399	0.484	0.429	0.492*	0.459	0.487	0.507	3.0%
Wiki	0.157	0.152	0.151	0.152	0.152	0.152	0.161*	0.158	0.152	0.139	0.172	0.176	9.3%
DBLP	0.443	0.443	0.443	0.443	0.448	0.443	0.447	0.517*	0.442	0.449	0.690	0.698	24.6%

Table 4: Overall performance comparison on classification

\uparrow

Datesets	BAED-enabled gradient-based explanation models							Non-GNN-based explanation algorithms					Improve.
Datesets	SM	inpgrad	IG	pgexp	gnnexp	gbp	deconv	GE-L	LIME	DW	LN-L	LP	Improve.
Cora	0.016	0.010	0.010	0.018	0.018	0.014	0.015	0.015*	0.060	0.017	0.017	-	33.3%
Citeseer	0.046	0.039	0.041	0.040	0.042	0.030	0.031	0.030*	0.274	0.041	0.042	0.045	0.0%
Pubmed	0.273	0.264	0.258	0.393	0.289	0.268	0.261	0.287	0.611	0.384	0.376	0.266*	3.0%
Wiki	1.178	1.178	1.164	1.179	1.177	1.157	1.139	1.149*	1.489	1.180	1.180	-	0.9%
DBLP	1.878	1.799	1.825	2.615	2.050	1.851	1.848	2.135*	3.600	2.612	2.634	2.192	15.7%

Table 5: Overall performance comparison on explainability

\downarrow

Datasets	Number of epochs per second $\uparrow$					BP time in BAED (sec.)	Total training time in seconds $\downarrow$					Improve.
Datasets	BAED	GCN	SAGE	tsGCN	DCI	BP time in BAED (sec.)	BAED	GCN	SAGE	tsGCN	DCI	Improve.
Cora	19.46	4.63	1.88	2.81	5.03	1.13	6.1	58.3	76.5	198.0	29.1	78.9%
Citeseer	10.10	5.47	3.40	3.28	3.81	0.11	10.1	31.9	30.3	159.2	32.0	66.3%
PubMed	2.60	2.21	1.66	0.84	2.82	0.06	38.1	51.6	60.0	827.0	653.1	26.8%
Wiki	3.65	3.43	2.21	1.98	5.03	0.15	27.1	29.4	45.2	268.1	32.4	6.4%
DBLP	1.67	1.94	2.60	1.00	1.19	1.11	61.1	50.8	38.7	716.8	218.9	-56.7%

Table 6: Overall performance comparison on efficiency

Datasets	Ablation-1	Ablation-2	Ablation-3	BAED
Datasets	B AED	BA ED	B A ED	BAED
Cora	16.70	30.15	30.15	51.81
Citeseer	8.97	18.16	20.21	78.67
PubMed	47.46	39.39	40.00	50.67
Wiki	8.10	16.12	15.72	12.71
DBLP	55.09	44.26	44.26	69.75

Table 7: Classification Accuracy (%) of Ablation Models

Figure 4: Classification accuracy with different labeling ratios

Size of Explanatory Subgraph	2	3	4	5	6	7	8
Cora	50.4	50.8	51.1	51.8	51.7	49.9	49.7
Citeseer	77.3	78.1	78.1	78.7	78.8	79.5	80.0
Pubmed	46.7	48.3	49.7	50.7	51.0	51.3	51.5
Wiki	12.1	12.3	12.5	12.7	12.9	13.1	13.3
dblp	68.9	68.9	69.8	69.8	70.1	70.1	70.1
Wisconsin	26.1	27.8	29.8	42.0	51.8	51.0	32.7

Table 8: Classification Accuracy (%) with Various Sizes of Explanatory Subgraphs

$\epsilon$	0.60	0.65	0.70	0.75	0.80	0.85	0.90	0.95	1.00
accuracy $\uparrow$	0.549	0.565	0.582	0.517	0.500	0.529	0.518	0.483	0.582
faithfulness $\downarrow$	0.004	0.004	0.004	0.004	0.004	0.005	0.016	0.009	0.010

Table 9: Prediction (accuracy) and explanation (faithfulness) performance of BAED under various

\epsilon

on Cora dataset.

Figure 5: Convergence Curves of BP Process in Label Augmentation. Legend denotes [dataset, No. converged iteration]

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Classification accuracy with different labeling ratios

Datasets	ratio(%)	GNNs				FSGL Models				BAED
Datasets	ratio(%)	SAGE	GAT	GIN	GCN	SGC	DLRGAE	tsGCN	DCI	SAGE+IG	SAGE+SM
Cora	2	0.309	0.193	0.120	0.348*	0.345	0.303	0.237	0.296	0.539	0.603
Cora	3	0.317	0.207	0.284	0.393	0.397*	0.285	0.282	0.348	0.616	0.546
Citeseer	2	0.180	0.242	0.219	0.241	0.247	0.221	0.273*	0.254	0.792	0.779
Citeseer	3	0.213	0.240	0.266	0.246	0.255	0.199	0.267	0.272*	0.712	0.732
PubMed	2	0.392	0.392	0.406	0.392	0.584*	0.478	0.549	0.523	0.632	0.618

Appendix

Table 10: Classification performance with different labeling ratios

Datasets	Ratio (%)	BAED-compatible gradient-based explanation models							Non-GNN-based explanation algorithms
Datasets	Ratio (%)	SM	inpgrad	IG	pgexp	gnnexp	gbp	deconv	GE-L	LIME	DW	LN-L	LP
Cora	2	0.0200	0.0249	0.0246	0.0222	0.0220	0.0241	0.0238	0.0143*	0.0710	0.0236	0.0226	-
Cora	3	0.0164	0.0098	0.0099	0.0183	0.0180	0.0143	0.0155	0.0153*	0.0601	0.0172	0.0174	-
Citeseer	2	0.0145	0.0148	0.0130	0.0205	0.0174	0.0136	0.0136	0.0129*	0.1466	0.0172	0.0176	0.0175
Citeseer	3	0.0457	0.0392	0.0411	0.0403	0.0416	0.0304	0.0313	0.0302	0.2739*	0.0405	0.0416	0.0453
Pubmed	2	0.2512	0.2442	0.2507	0.3418	0.2497	0.2415	0.2332	0.2674	0.6142	0.3371	0.3307	0.2412*

Appendix

Table 11: Explainability performance with different labeling ratios

Figure 7: Convergence Curves of BP Process in Label Augmentation. Labeling Ratio = 2% (Left) and 3% (Right).

Graph-Based Label Propagation

See all 16 papers

Advancing Graph Few-Shot Learning via In-Context Learning
May 23, 2026Renchu Guan, Yajun Wang, Chunli Guo +5Meta-Learning Few-Shot Learning
Key Laboratory of Symbolic Computation and Knowledge Engineering of the Ministry of Education
Improving Graph Few-shot Learning with Hyperbolic Space and Denoising Diffusion
Apr 30, 2026Yonghao Liu, Jialu Sun, Wei Pang +4Few-Shot Learning
Key Laboratory of Symbolic Computation and Knowledge Engineering of the Ministry of Education, College of Computer Science and Technology, Jilin University · School of Mathematical and Computer Sciences, Heriot-Watt University · Department of Information Engineering and Computer Science, University of Trento
Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance
Aug 14, 2026Taraneh Younesian, Steve Azzolin, Antonio Longa +3Graph Neural Networks Shortcut Learning
VU Amsterdam, Netherlands · University of Trento, Italy · UiT The Arctic University of Norway, Norway

May 23, 2026J/K move · Enter open · S save

Advancing Graph Few-Shot Learning via In-Context Learning

Renchu Guan, Yajun Wang, Chunli Guo, Bowen Cao+4

Key Laboratory of Symbolic Computation and Knowledge Engineering of the Ministry of Education

Meta-Learning Few-Shot Learning In-Context Learning

Open paper AskPDF arXiv

Graph few-shot learning, which aims to classify nodes from novel classes with only a few labeled examples, is a widely studied problem in graph learning. However, existing methods often face two key limitations. First, the predominant graph few-shot learning paradigm relies on supervised tasks, failing to leverage the vast number of unlabeled nodes in the graph. Second, many approaches require complex task adaptation or fine-tuning during inference, limiting their efficiency and applicability. Inspired by the powerful in-context learning capabilities of large language models, we propose a novel model named VISION for adVancIng graph few-Shot learning via In-cOntext LearNing to address these challenges. Our model reframes graph few-shot learning as a fine-tuning-free sequence reasoning problem. At its core is a context-aware network that initializes nodes with role embeddings and employs a dual-context fusion module to synergistically integrate local topological structures and global task-level dependencies. This allows our model to dynamically generate class-aware representations for the query set conditioned on the support set context in a single forward pass. To effectively train our model, we introduce an unsupervised task generator that creates structure-adaptive features and constructs diverse pseudo-tasks from abundant unlabeled data. Our method unifies unsupervised meta-learning with graph in-context learning, achieving efficient inference. Extensive experiments on multiple benchmark datasets demonstrate the superiority of our model. Our public code can be found

Abstract

Graph-Based Label Propagation

Explore similar work

Advancing Graph Few-Shot Learning via In-Context Learning

Improving Graph Few-shot Learning with Hyperbolic Space and Denoising Diffusion

Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance