Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance

Taraneh Younesian; Steve Azzolin; Antonio Longa; Francesco Ferrini; Vincenzo Marco De Luca; Stefano Teso

cs.LGAug 14, 2026

Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance

Authors: Taraneh Younesian, Steve Azzolin, Antonio Longa, Francesco Ferrini, Vincenzo Marco De Luca, Stefano Teso

Organizations: VU Amsterdam, Netherlands · University of Trento, Italy · UiT The Arctic University of Norway, Norway

Abstract

Graph Neural Networks (GNNs) can solve prediction tasks by unintentionally exploiting shortcuts---that is, edges, nodes, and features that correlate with but are not causal for the prediction---which compromise their reliability in out-of-distribution tasks. We introduce XIGL, an architecture-agnostic human-in-the-loop strategy for removing such shortcuts from GNNs. Our key insight is twofold. On the one hand, reliance on shortcuts can be detected by inspecting GNN explanations. On the other hand, once made aware of such shortcuts, sufficiently expert users can provide tailored corrective feedback, which helps deconfound the model. XIGL supports any query strategy; however, since corrective feedback can be expensive to acquire, we develop an active learning strategy for prioritizing explanations that are more likely to display shortcut behavior, lowering annotation and cognitive costs. We showcase the effectiveness of XIGL, including both existing and proposed explanation-based strategies, on several GNN architectures. Our implementation is available online.

Figures & tables

Figure 1: Overview of XIGL . We first extract explanations from the GNN being trained over a small subset of the dataset; an active learning strategy then selects the most informative explanations for an expert to correct by indicating which nodes the model should not rely on and which nodes it should rely on instead; the corrections then fine-tune the GNN via an explanation-guided loss that removes shortcut dependence. The process iterates until the annotation budget is reached.

\mathcal{D}_{\mathrm{exp}}=\{(G_{i},y_{i},E_{i})\}_{i=1}^{q}.

(2)

Algorithm 1 Active Learning Steps in XIGL

Panel of Figure 2 — Figure 2: ER-color dataset examples with and without confounders. The trees shown are the ground-truth explanations for each class. As confounders, for $y=0$ , a purple node is added to the graph, and for $y=1$ , a cyan node is added.

	ER-color		CPatchMNIST
	No Sup.	Passive	No Sup.	Passive
GCN	$60.40{\scriptstyle\pm 11.36}$	$\bm{71.20}{\scriptstyle\pm 20.73}$	$26.28{\scriptstyle\pm 12.58}$	$\bm{30.57}{\scriptstyle\pm 2.90}$
GIN	$52.93{\scriptstyle\pm 4.23}$	$\bm{74.20}{\scriptstyle\pm 21.18}$	$0.20{\scriptstyle\pm 0.12}$	$\bm{33.38}{\scriptstyle\pm 1.13}$
GraphSAGE	$56.80{\scriptstyle\pm 14.24}$	$\bm{70.53}{\scriptstyle\pm 17.76}$	$\bm{31.5}{\scriptstyle\pm 6.85}$	$30.68{\scriptstyle\pm 1.88}$
GAT	$50.93{\scriptstyle\pm 4.21}$	$\bm{74.27}{\scriptstyle\pm 18.74}$	$22.61{\scriptstyle\pm 3.93}$	$\bm{25.25}{\scriptstyle\pm 9.38}$

Table 1: Effect of passive supervision in XIGL across datasets. Accuracy (%) on the unconfounded test split. Results are averaged with standard deviation across 5 seeds.

	ER-color				CPatchMNIST
	Random	MaCE	MaEE	MiEE	Random	MaCE	MaEE	MiEE
GCN	$58.13{\scriptstyle\pm 21.07}$	$54.67{\scriptstyle\pm 9.98}$	$59.87{\scriptstyle\pm 20.43}$	$\bm{64.00}{\scriptstyle\pm 13.93}$	$26.52{\scriptstyle\pm 3.48}$	$22.56{\scriptstyle\pm 6.21}$	$28.21{\scriptstyle\pm 2.10}$	$\bm{28.75}{\scriptstyle\pm 0.73}$
GIN	$\bm{68.80}{\scriptstyle\pm 22.93}$	$63.87{\scriptstyle\pm 12.40}$	$66.67{\scriptstyle\pm 27.13}$	$66.93{\scriptstyle\pm 19.50}$	$\bm{32.74}{\scriptstyle\pm 2.80}$	$25.50{\scriptstyle\pm 3.88}$	$30.87{\scriptstyle\pm 2.61}$	$27.77{\scriptstyle\pm 2.77}$
GraphSAGE	$59.87{\scriptstyle\pm 16.67}$	$\bm{73.20}{\scriptstyle\pm 23.74}$	$55.20{\scriptstyle\pm 7.49}$	$56.27{\scriptstyle\pm 11.90}$	$25.86{\scriptstyle\pm 2.30}$	$22.57{\scriptstyle\pm 8.46}$	$26.38{\scriptstyle\pm 5.25}$	$\bm{28.28}{\scriptstyle\pm 3.94}$
GAT	$52.67{\scriptstyle\pm 14.12}$	$\bm{60.40}{\scriptstyle\pm 22.10}$	$53.87{\scriptstyle\pm 9.81}$	$50.13{\scriptstyle\pm 1.79}$	$26.90{\scriptstyle\pm 2.34}$	$26.59{\scriptstyle\pm 1.22}$	$\bm{29.20}{\scriptstyle\pm 11.64}$	$28.57{\scriptstyle\pm 5.86}$

Table 2: Effect of AL strategies of XIGL across datasets. Accuracy (%) on the unconfounded test split. Results are averaged with standard deviation across 5 seeds.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

	ER-color								CPatchMNIST
	# Layers	H	Dropout	Aggr.	LR	$\lambda$	Start E.	# E	# Layers	H	Dropout	Aggr.	LR	$\lambda$	Start E.	# E
GCN	3	100	0.3	add	$9e-7$	10	100	400	3	128	0.3	add	$3e-6$	$19.22$	30	500
GIN	3	64	0	mean	$5e-4$	10	50	300	5	256	0	mean	$8e-7$	$38.77$	30	300
SAGE	3	128	0	add	$6e-6$	1	30	300	5	128	0	add	$3e-6$	$71.77$	70	500
GAT	3	128	0	add	$3.e-5$	1	0	300	3	64	0	mean	$5e-5$	$22.56$	0	200

Appendix

Table 3: Hyperparameter settings for different GNN architectures for the passive supervision setup on ER-color and CPatchMNIST . “H” indicated the hidden dimension, “Aggr.” indicated the aggregation function of the GNN, “Start E.” indicates the starting epoch for training with both losses, and “E”. indicated the total training epochs.

	ER-color								CPatchMNIST
	Random		MaCE		MaEE		MiEE		Random		MaCE		MaEE		MiEE
	LR	$\lambda$	LR	$\lambda$	LR	$\lambda$	LR	$\lambda$	LR	$\lambda$	LR	$\lambda$	LR	$\lambda$	LR	$\lambda$
GCN	$1e-5$	$1.09$	$3e-5$	$1.19$	$7e-6$	$26.40$	$5e-5$	$25.21$	$6e-7$	$251.75$	$6e-6$	$614.48$	$6e-7$	$317.17$	$8e-7$	$446.70$
GIN	$2e-4$	$29.91$	$7e-4$	$1.06$	$6e-4$	$81.08$	$2e-4$	$2.51$	$6e-7$	$162.74$	$2e-6$	$672.56$	$3e-6$	$606.85$	$6e-7$	$212.88$
SAGE	$5e-5$	$23.91$	$1e-4$	$20.10$	$5e-5$	$22.23$	$8e-5$	$57.29$	$9e-6$	$325.70$	$3e-5$	$380.74$	$2e-5$	$850.60$	$3e-6$	$102.23$
GAT	$9e-5$	$751.34$	$6e-5$	$565.93$	$5e-5$	$565.52$	$9e-6$	$247.76$	$4e-5$	$929.88$	$2e-5$	$130.81$	$7e-5$	$482.22$	$1e-4$	$961.77$

Appendix

Table 4: Training hyperparameters for different GNN architectures on ER-color and CPatchMNIST .

Graph Neural Networks

Graph Neural Networks learn from graph-structured data by combining information across connected nodes, edges, and larger structures.

See all 701 papers

Declarative Explanations for Graph Neural Networks: A Demonstration
Oct 19, 2026Cibo Yu, Haitong Tang, Yinghui Wu +4GNN Explainability Graph Neural Networks
Zhejiang University, Hangzhou, China · Case Western Reserve University, Cleveland, USA · Bowling Green State University, Bowling Green, USA
Why Self-Inconsistency Arises in GNN Explanations and How to Exploit It
May 8, 2026Wenxin Tai, Yaqian Liu, Ting Zhong +1Explanation Stability GNN Explainability
University of Electronic Science and Technology of China
A Completion-Aware Framework for Impactful Counterfactual Explainability in Graph Neural Networks
Jun 20, 2026Maria Myrto Villia, Filippos Gouidis, Theodore Patkos +1Counterfactual Explanations GNN Explainability
Institute of Computer Science, Foundation for Research and Technology - Hellas (FORTH), Heraklion, Greece · Computer Science Department, University of Crete, Greece

Oct 19, 2026J/K move · Enter open · S save

Declarative Explanations for Graph Neural Networks: A Demonstration

Cibo Yu, Haitong Tang, Yinghui Wu, Arijit Khan+3

Zhejiang University, Hangzhou, China · Case Western Reserve University, Cleveland, USA · Bowling Green State University, Bowling Green, USA

GNN Explainability Graph Neural Networks

Open paper AskPDF arXiv

Graph Neural Networks (GNNs) have proven effective for graph analytical tasks, yet their "black-box" nature complicates the understanding of their layer-wise behavior and hinders their application to large graphs. We demonstrate SliceGXQ, an end-to-end declarative and interactive system for querying and exploring layer-wise explanations for GNNs over large-scale graphs. In particular, SliceGXQ provides (1) an interactive graphical user interface for dataset, model, and explanation configuration; (2) a lightweight SPARQL-like parser that allows users to specify target nodes, layers, explanation modes, and constraints; and (3) an integrated execution and visualization workflow that invokes SliceGX as its underlying explanation engine. We demonstrate how users can diagnose a misclassified node, identify problematic layers and influential messages, and perform targeted model debugging, together with large-scale graph interpretation and fraud-detection analysis. Our demonstration video and source code are available at https://github.com/jinzhaoqet/SliceGXQ.

Abstract

Graph Neural Networks

Explore similar work

Declarative Explanations for Graph Neural Networks: A Demonstration

Why Self-Inconsistency Arises in GNN Explanations and How to Exploit It

A Completion-Aware Framework for Impactful Counterfactual Explainability in Graph Neural Networks