GNN-VIEW: Visual Inspection and Explanation of Graph-based Earth-System Models

Alessandro Lucatelli; Ferdinand Unterhuber; Katharina Krenzer; Mennatallah El-Assady; Benedikt Soja; Christina Humer; Joel Oskarsson

cs.LGOct 26, 2026

GNN-VIEW: Visual Inspection and Explanation of Graph-based Earth-System Models

Authors: Alessandro Lucatelli, Ferdinand Unterhuber, Katharina Krenzer, Mennatallah El-Assady, Benedikt Soja, Christina Humer, Joel Oskarsson

Organizations: ETH Zürich, Switzerland

Abstract

Machine learning Earth-system models are becoming increasingly important, and so is their development. Graph-based models, one of the most prominent approaches, allow developers to constrain the learning process before training via graph connectivity. Yet, developing these models remains challenging: the graph is built algorithmically with little intuition about the final result beyond static plotting, which requires extra work and is often not detailed enough. Even after successful training, how the model arrives at forecasts remains obscure. We present GNN-VIEW, an interactive browser-based tool that supports the development of such systems. GNN-VIEW exposes the graph's information pathways, allowing model developers to debug the graph structure before training. After training, GNN-VIEW uses attribution-based explainability to reveal which input features are used by the trained model for a given forecast. This helps developers understand why a model fails. We showcase the usefulness of GNN-VIEW in the construction stage of a graph for an ocean forecasting model, and in explaining poorly predicted precipitation of a trained weather forecasting model during Hurricane Laura.

Figures & tables

Figure 1: GNN-VIEW follows a developer’s workflow. ( a ) Before training, it allows ( b ) exploration of the graph structure and its components (grid points, mesh nodes, edges), followed by ( c ) detailed inspection of the graph structure via ( d-e ) manual thresholding of graph statistics. ( f ) After model training, developers can get ( i-k ) attribution-based explanations for ( g,h ) predictions of interest to detect what lies behind model failure.

Figure 2: GNN-VIEW Inspect mode, on the Njord ocean graph. ( a-b ) Grid points thresholded on out-degree in the encoder subgraph (G2M). ( c-e ) Mesh edges thresholded on length and mesh nodes thresholded on degree. All distributions span the full graph, not just the viewport in the figure.

Figure 3: GNN-VIEW Explain mode on a GraphCast-style GNN. ( Black square ) Target node 31.5° N 93.0° W. XAI target field is ( Left ) Temperature at 925 hPa and ( Right ) Total 6-hourly Precipitation. ( a,c ) Target error: absolute standardized error heatmap of the target field. ( b,d ) Input attribution: signed IG attribution of the target node’s prediction w.r.t. that field at t $-$ 6 h, scaled to each panel’s own largest absolute attribution excluding the target node. ( e,f ) Ranked input fields: top-10 input features by IG attribution share.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Server latency measured for ( a-c ) one component at a time, and ( d ) for real graphs. ( a ) Latency of rendering an increasing number of grid points; mesh nodes and edges count kept fixed. ( b ) Latency of rendering an increasing number of mesh nodes; grid-points and edge count kept fixed. ( c ) Latency of rendering an increasing number of edges; grid-point and mesh-node count kept fixed. ( d ) Latency of rendering the default graphs available in GNN-VIEW: GraphCast-style and Graph-EFM graphs for weather modeling from [ 24 , 28 ] and SeaCast-style and Njord graphs for ocean modeling from [ 18 , 19 ] .

Figure 5: Randomization tests for Hurricane Laura use case in Sec. 3 . Target XAI field is ( Left ) Temperature at 925 hPa, ( Right ) Precipitation. ( a,c ) Similarity of IG input attribution maps for the 4 paired tests described in App. C.1 measured as Pearson’s correlation. ( b,d ) Similarity of IG rankings measured as Kendall’s correlation. ( Second row ) IG explanations for the Laura use case. ( Third row ) IG explanations for a model with randomized weights. For each target the seed whose IG attribution has the highest correlation with the trained model’s IG attribution is shown. ( Fourth row ) Model inputs for the target’s own field at t

-

6 h. Correlations with the corresponding entries in the Second row are reported.

Figure 6: Randomization test for ( Left ) 50 random targets (S0), and ( Right ) 50 targets with highest prediction error (S1). ( First row ) IG. ( Second row ) IxG. ( Brackets ) Two-sided Wilcoxon signed-rank test, paired by target (ns

>0.05

, *

\leq 0.05

, **

\leq 0.01

, ***

\leq 10^{-3}

, ****

\leq 10^{-4}

p

-values are Bonferroni-corrected.

Figure 7: Deletion tests for the Hurricane Laura use case in Sec. 3 . Target XAI field is ( Left ) Temperature at 925 hPa, ( Right ) Precipitation. ( a,c ) Perturbation curves when deleting single entries, ranked by (green) trained model’s IG attributions, (gray) input magnitude, and (light gray) ranked at random. ( b,d ) Perturbation curves when deleting whole input fields, ranked by (green) trained model’s IG attribution share, (gray) input field magnitude share, and (light gray) at random. ( e,f ) Faithfulness correlations.

Figure 8: Deletion tests for ( Left ) 50 random targets (S0), and ( Right ) 50 targets with highest prediction error (S1). ( First row ) IG attributions ( Second row ) IxG. ( a,c,e,g ) Perturbation curves when deleting single entries, ranked by (green) trained model’s IG/IxG attributions, (gray) input magnitude, and (light gray) ranked at random. ( b,d,f,h ) Perturbation curves when deleting whole input fields, ranked by (green) trained model’s IG/IxG attribution share, (gray) input field magnitude share, and (light gray) at random. Lines are medians over targets, shaded areas show the range from the 25th to the 75th percentiles. ( Third row ) Faithfulness correlations for ( i,k ) IG and ( j,l ) IxG. ( Brackets ) Two-sided Wilcoxon signed-rank test of trained vs. input-only

\rho

, paired by target (ns

>0.05

, *

\leq 0.05

, **

\leq 0.01

, ***

\leq 10^{-3}

, ****

\leq 10^{-4}

p

-values are Bonferroni-corrected.

Figure 9: IG completeness tests for ( First row ) 50 random targets (S0), and ( Second row ) 50 targets with highest prediction error (S1). ( a,d ) Convergence of IG input attribution maps. ( b,e ) Convergence of IG input rankings. ( c,f ) Completeness gap with increasing number of integration steps (

m

Figure 10: IG completeness tests for the Hurricane Laura use case in Sec. 3 . Target XAI field is ( Left ) Temperature at 925 hPa, ( Right ) Precipitation. ( First row ) IG explanations for the Laura use case computed with $m=64$ integration steps. ( e,g ) Convergence of IG input attribution maps. ( f,h ) Convergence of IG input rankings. In all four plots the completeness gap for using a given $m$ is reported on the horizontal axis. The vertical axis reports correlations between attributions computed with $m$ and $2m$ integration steps.

GNN Explainability

Graph neural network explainability studies how to identify and communicate the features, nodes, edges, and subgraphs behind graph-model predictions.

See all 82 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
Beyond Trained Models: Compiling GNNs for a Sound Explainer Benchmark
Oct 2, 2026Steve Azzolin, Francesco Paolo Nerini, Stefano Teso +4GNN Explainability Explainability Evaluation
University of Trento Trento, Italy
GRAFT: Auditing Graph Neural Networks via Global Feature Attribution
May 5, 2026Rishi Raj Sahoo, Subhankar MishraFeature Attribution Graph Neural Networks
National Institute of Science Education and Research (NISER), Bhubaneswar, India · Homi Bhabha National Institute, Mumbai, India

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

GNN Explainability

Explore similar work

Declarative Explanations for Graph Neural Networks: A Demonstration

Beyond Trained Models: Compiling GNNs for a Sound Explainer Benchmark

GRAFT: Auditing Graph Neural Networks via Global Feature Attribution