Relevant Walk Search for Explaining Graph Neural Networks
Pith reviewed 2026-05-25 05:09 UTC · model grok-4.3
The pith
Polynomial-time max-product algorithms identify top-K relevant walks for GNN explanations exactly at the neuron level.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The central claim is that algorithms derived from the max-product method can locate the top-K relevant walks in polynomial time, exactly at the neuron level and approximately at the node level, which allows GNN-LRP to scale to large problems while maintaining its higher-order explanatory power.
What carries the argument
The max-product algorithm, which finds maximum likelihood configurations in graphical models, adapted here to propagate and select the highest relevance scores along walks in the GNN.
If this is right
- Scales GNN-LRP explanations to graphs too large for exhaustive walk search.
- Preserves exact top-K walks at neuron level without losing higher-order information.
- Provides practical approximations at node level for broader applicability.
- Validated on benchmarks from epidemiology, molecular chemistry, and natural language processing.
Where Pith is reading between the lines
- The method may generalize to other path-based explanation techniques in neural networks.
- Could support interactive explanation tools for GNNs in production settings.
- Opens the way for combining with sampling methods for even larger networks.
Load-bearing premise
The max-product algorithm can be directly adapted to GNN-LRP relevance scores to find the exact top-K walks at the neuron level.
What would settle it
Compare the top-K walks found by the new algorithm against those from full enumeration on a small-depth GNN where exponential computation is still feasible.
Figures
read the original abstract
Graph Neural Networks (GNNs) have become important machine learning tools for graph analysis, and its explainability is crucial for safety, fairness, and robustness. Layer-wise relevance propagation for GNNs (GNN-LRP) evaluates the relevance of \emph{walks} to reveal important information flows in the network, and provides higher-order explanations, which have been shown to be superior to the lower-order, i.e., node-/edge-level, explanations. However, identifying relevant walks by GNN-LRP requires {\em exponential} computational complexity with respect to the network depth, which we will remedy in this paper. Specifically, we propose {\em polynomial-time} algorithms for finding top-$K$ relevant walks, which drastically reduces the computation and thus increases the applicability of GNN-LRP to large-scale problems. Our proposed algorithms are based on the \emph{max-product} algorithm -- a common tool for finding the maximum likelihood configurations in probabilistic graphical models -- and can find the most relevant walks exactly at the neuron level and approximately at the node level. Our experiments demonstrate the performance of our algorithms at scale and their utility across application domains, i.e., on epidemiology, molecular, and natural language benchmarks. We provide our codes under \href{https://github.com/xiong-ping/rel_walk_gnnlrp}{github.com/xiong-ping/rel\_walk\_gnnlrp}.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes polynomial-time algorithms, based on the max-product algorithm, for identifying the top-K most relevant walks under GNN-LRP. These algorithms are claimed to recover the walks exactly at the neuron level and approximately at the node level, thereby replacing the exponential enumeration required by standard GNN-LRP with scalable dynamic programming. Experiments on epidemiology, molecular, and NLP benchmarks are presented to demonstrate runtime and utility, and code is released.
Significance. If the exactness claim at the neuron level holds, the work would materially increase the applicability of higher-order walk-based explanations to deeper and larger GNNs, a recognized bottleneck in current GNN explainability. The public code release is a concrete strength that supports reproducibility and follow-up work.
major comments (2)
- [§3 (algorithm derivation)] The central claim of exact top-K recovery at the neuron level (abstract and §3) rests on repurposing max-product without introducing unstated approximation errors. The manuscript must explicitly show that GNN-LRP relevance scores satisfy the non-negativity and independence conditions required for the dynamic program to match exhaustive enumeration; otherwise the polynomial algorithm may return a different set than the exponential baseline.
- [Table 2 / experimental verification] Table 2 or the corresponding runtime table: the reported speed-ups are only meaningful if the returned walks are verified to be identical (neuron level) or within a stated bound (node level) to the exhaustive top-K; without such a verification column the empirical results do not yet confirm the exactness property.
minor comments (2)
- [§2 / §3] Notation for neuron-level versus node-level walks is introduced without a clear running example; a small illustrative graph with explicit relevance scores would clarify the distinction.
- [Abstract / §3] The abstract states the algorithms are 'based on the max-product algorithm' but does not cite the specific PGM reference or the precise variant (e.g., with or without beam pruning) used for the top-K extension.
Simulated Author's Rebuttal
We thank the referee for the careful reading and constructive comments on the exactness and empirical verification of our algorithms. We address each major comment below and will incorporate revisions to strengthen the manuscript.
read point-by-point responses
-
Referee: [§3 (algorithm derivation)] The central claim of exact top-K recovery at the neuron level (abstract and §3) rests on repurposing max-product without introducing unstated approximation errors. The manuscript must explicitly show that GNN-LRP relevance scores satisfy the non-negativity and independence conditions required for the dynamic program to match exhaustive enumeration; otherwise the polynomial algorithm may return a different set than the exponential baseline.
Authors: We agree that an explicit derivation of the required conditions is needed for full rigor. In the revised version we will insert a short subsection in §3 that proves GNN-LRP relevance scores are non-negative (by construction of the standard LRP rules) and that the additive decomposition of walk relevance satisfies the independence property needed for the max-product dynamic program to recover exactly the same top-K set as exhaustive enumeration at the neuron level. revision: yes
-
Referee: [Table 2 / experimental verification] Table 2 or the corresponding runtime table: the reported speed-ups are only meaningful if the returned walks are verified to be identical (neuron level) or within a stated bound (node level) to the exhaustive top-K; without such a verification column the empirical results do not yet confirm the exactness property.
Authors: We accept this observation. We will augment Table 2 with an additional verification column that reports (i) exact match rate with exhaustive enumeration on all neuron-level instances small enough for brute-force comparison, and (ii) the empirical deviation from the stated approximation bound on the node-level instances. This will directly confirm the claimed exactness and approximation guarantees. revision: yes
Circularity Check
No circularity: algorithmic reduction is independent of inputs
full rationale
The paper's core contribution is a new polynomial-time algorithm adapting the max-product method from PGMs to compute top-K relevant walks for GNN-LRP, claimed to be exact at the neuron level. This is presented as an independent algorithmic engineering step that reduces exponential enumeration to polynomial time without any fitted parameters, self-definitional mappings, or load-bearing self-citations in the abstract or claimed derivation. No equations or results reduce by construction to the input relevance scores; the exactness claim is an asserted property of the DP adaptation rather than a renaming or re-expression of existing quantities. The derivation chain is therefore self-contained against external algorithmic benchmarks.
Axiom & Free-Parameter Ledger
Reference graph
Works this paper leans on
-
[1]
Zonghan Wu and Shirui Pan and Fengwen Chen and Guodong Long and Chengqi Zhang and Philip S. Yu , title =
-
[2]
On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation , author=. PloS one , volume=. 2015 , publisher=
work page 2015
-
[3]
Complex network measures of brain connectivity: uses and interpretations , author=. Neuroimage , volume=. 2010 , publisher=
work page 2010
-
[4]
and Chmiela, Stefan and Gastegger, Michael and Schütt, Kristof T
Unke, Oliver T. and Chmiela, Stefan and Gastegger, Michael and Schütt, Kristof T. and Sauceda, Huziel E. and Müller, Klaus-Robert , doi =. SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects , ty =. Nature Communications , number =
-
[5]
Human cognition involves the dynamic integration of neural activity and neuromodulatory systems , author=. Nature neuroscience , volume=. 2019 , publisher=
work page 2019
-
[6]
Digital Signal Processing , volume=
Methods for interpreting and understanding deep neural networks , author=. Digital Signal Processing , volume=. 2018 , publisher=
work page 2018
-
[7]
Layer-Wise Relevance Propagation: An Overview , booktitle =
Gr. Layer-Wise Relevance Propagation: An Overview , booktitle =
-
[8]
Explaining Deep Neural Networks and Beyond:
Wojciech Samek and Gr. Explaining Deep Neural Networks and Beyond:. Proc
- [9]
-
[10]
Hao Yuan and Haiyang Yu and Shurui Gui and Shuiwang Ji , title =. CoRR , volume =. 2020 , url =
work page 2020
-
[11]
GNNExplainer: Generating Explanations for Graph Neural Networks , booktitle =
Zhitao Ying and Dylan Bourgeois and Jiaxuan You and Marinka Zitnik and Jure Leskovec , editor =. GNNExplainer: Generating Explanations for Graph Neural Networks , booktitle =
-
[12]
Parameterized Explainer for Graph Neural Network , booktitle =
Dongsheng Luo and Wei Cheng and Dongkuan Xu and Wenchao Yu and Bo Zong and Haifeng Chen and Xiang Zhang , editor =. Parameterized Explainer for Graph Neural Network , booktitle =
-
[13]
On Explainability of Graph Neural Networks via Subgraph Explorations , booktitle =
Hao Yuan and Haiyang Yu and Jie Wang and Kang Li and Shuiwang Ji , editor =. On Explainability of Graph Neural Networks via Subgraph Explorations , booktitle =
-
[14]
Minh N. Vu and My T. Thai , editor =. Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual , year =
work page 2020
-
[15]
7th International Conference on Learning Representations,
Keyulu Xu and Weihua Hu and Jure Leskovec and Stefanie Jegelka , title =. 7th International Conference on Learning Representations,
-
[16]
Thomas Schnake and Oliver Eberle and Jonas Lederer and Shinichi Nakajima and Kristof T. Sch. Higher-Order Explanations of Graph Neural Networks via Relevant Walks , journal =
-
[17]
Link Prediction Based on Graph Neural Networks , booktitle =
Muhan Zhang and Yixin Chen , editor =. Link Prediction Based on Graph Neural Networks , booktitle =
-
[18]
Explaining decisions of graph convolutional neural networks: patient-specific molecular subnetworks responsible for metastasis prediction in breast cancer , author=. Genome Medicine , volume=. 2021 , publisher=
work page 2021
-
[19]
6th International Conference on Learning Representations,
Jie Chen and Tengfei Ma and Cao Xiao , title =. 6th International Conference on Learning Representations,
-
[20]
Evolution of Graph Classifiers , year=
Domingue, Miguel and Dhamdhere, Rohan and Harish Kanamarlapudi, Naga Durga and Raghupathi, Sunand and Ptucha, Raymond , booktitle=. Evolution of Graph Classifiers , year=
-
[21]
Hamilton and Zhitao Ying and Jure Leskovec , editor =
William L. Hamilton and Zhitao Ying and Jure Leskovec , editor =. Inductive Representation Learning on Large Graphs , booktitle =
-
[22]
The Journal of Chemical Physics , volume=
Schnet--a deep learning architecture for molecules and materials , author=. The Journal of Chemical Physics , volume=. 2018 , publisher=
work page 2018
-
[23]
Franco Scarselli and Marco Gori and Ah Chung Tsoi and Markus Hagenbuchner and Gabriele Monfardini , title =
-
[24]
Kipf and Max Welling , title =
Thomas N. Kipf and Max Welling , title =. 5th International Conference on Learning Representations,
-
[25]
Olah, Chris and Mordvintsev, Alexander and Schubert, Ludwig , title =. Distill , year =
-
[26]
Karen Simonyan and Andrea Vedaldi and Andrew Zisserman , editor =. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps , booktitle =
-
[27]
Real Time Image Saliency for Black Box Classifiers , booktitle =
Piotr Dabkowski and Yarin Gal , editor =. Real Time Image Saliency for Black Box Classifiers , booktitle =
-
[28]
Jianbo Chen and Le Song and Martin J. Wainwright and Michael I. Jordan , editor =. Learning to Explain: An Information-Theoretic Perspective on Model Interpretation , booktitle =
-
[29]
Hao Yuan and Jiliang Tang and Xia Hu and Shuiwang Ji , editor =
-
[30]
Federico Baldassarre and Hossein Azizpour , title =. CoRR , volume =. 2019 , url =
work page 2019
-
[31]
Pope and Soheil Kolouri and Mohammad Rostami and Charles E
Phillip E. Pope and Soheil Kolouri and Mohammad Rostami and Charles E. Martin and Heiko Hoffmann , title =
-
[32]
Yue Zhang and David DeFazio and Arti Ramesh , editor =. RelEx:
-
[33]
Qiang Huang and Makoto Yamada and Yuan Tian and Dinesh Singh and Dawei Yin and Yi Chang , title =. CoRR , volume =. 2020 , url =
work page 2020
-
[34]
Spectral Networks and Locally Connected Networks on Graphs , booktitle =
Joan Bruna and Wojciech Zaremba and Arthur Szlam and Yann LeCun , editor =. Spectral Networks and Locally Connected Networks on Graphs , booktitle =
-
[35]
Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering , booktitle =
Micha. Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering , booktitle =
- [36]
-
[37]
and Debnath, Gargi and Shusterman, Alan J
Debnath, Asim Kumar and Lopez de Compadre, Rosa L. and Debnath, Gargi and Shusterman, Alan J. and Hansch, Corwin , title =. Journal of Medicinal Chemistry , volume =. 1991 , doi =
work page 1991
-
[38]
Journal of Medicinal Chemistry , volume =
Kazius, Jeroen and McGuire, Ross and Bursi, Roberta , title =. Journal of Medicinal Chemistry , volume =. 2005 , doi =
work page 2005
-
[39]
Richard Socher and Alex Perelygin and Jean Wu and Jason Chuang and Christopher D. Manning and Andrew Y. Ng and Christopher Potts , title =. Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing,
work page 2013
-
[40]
Pinar Yanardag and S. V. N. Vishwanathan , editor =. Deep Graph Kernels , booktitle =. 2015 , doi =
work page 2015
-
[41]
arXiv preprint arXiv:2005.00687 , year=
Open Graph Benchmark: Datasets for Machine Learning on Graphs , author=. arXiv preprint arXiv:2005.00687 , year=
-
[42]
A Unified Approach to Interpreting Model Predictions , volume =
Lundberg, Scott M and Lee, Su-In , booktitle =. A Unified Approach to Interpreting Model Predictions , volume =
-
[43]
Gilmer, Justin and Schoenholz, Samuel S. and Riley, Patrick F. and Vinyals, Oriol and Dahl, George E. , title =. Proceedings of the 34th International Conference on Machine Learning - Volume 70 , pages =. 2017 , publisher =
work page 2017
-
[44]
Journal of Medicinal Chemistry , volume =
Kazius, Jeroen and McGuire, Ross and Bursi, Roberta , title =. Journal of Medicinal Chemistry , volume =
-
[45]
Yuyang Gao and Tong Sun and Rishab Bhatt and Dazhou Yu and Sungsoo Hong and Liang Zhao , editor =
-
[46]
Efficient Computation of Higher-Order Subgraph Attribution via Message Passing , booktitle =
Ping Xiong and Thomas Schnake and Gr. Efficient Computation of Higher-Order Subgraph Attribution via Message Passing , booktitle =. 2022 , url =
work page 2022
-
[47]
Statistics and computing , volume=
An efficient algorithm for finding the m most probable configurationsin probabilistic expert systems , author=. Statistics and computing , volume=. 1998 , publisher=
work page 1998
-
[48]
Optimizing sentinel surveillance in temporal network epidemiology , author=. Scientific reports , volume=. 2017 , publisher=
work page 2017
-
[49]
Kriege and Christopher Morris and Petra Mutzel , editor =
Lutz Oettershagen and Nils M. Kriege and Christopher Morris and Petra Mutzel , editor =. Temporal Graph Kernels for Classifying Dissemination Processes , booktitle =. 2020 , url =. doi:10.1137/1.9781611976236.56 , timestamp =
-
[50]
What's in a crowd? Analysis of face-to-face behavioral networks , journal =
Lorenzo Isella and Juliette Stehlé and Alain Barrat and Ciro Cattuto and Jean-François Pinton and Wouter. What's in a crowd? Analysis of face-to-face behavioral networks , journal =. 2011 , issn =. doi:https://doi.org/10.1016/j.jtbi.2010.11.033 , url =
-
[51]
Moghaddam, Amin and Wattenhofer, Roger , title =
Faber, Lukas and K. Moghaddam, Amin and Wattenhofer, Roger , title =. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , pages =. 2021 , isbn =. doi:10.1145/3447548.3467283 , abstract =
-
[52]
Axiomatic Attribution for Deep Networks , booktitle =
Mukund Sundararajan and Ankur Taly and Qiqi Yan , editor =. Axiomatic Attribution for Deep Networks , booktitle =. 2017 , url =
work page 2017
-
[53]
Hamilton and Zhitao Ying and Jure Leskovec , editor =
William L. Hamilton and Zhitao Ying and Jure Leskovec , editor =. Inductive Representation Learning on Large Graphs , booktitle =. 2017 , url =
work page 2017
-
[54]
Zonghan Wu and Shirui Pan and Fengwen Chen and Guodong Long and Chengqi Zhang and Philip S. Yu , title =. 2021 , url =. doi:10.1109/TNNLS.2020.2978386 , timestamp =
-
[55]
Fabian Yamaguchi and Nico Golde and Daniel Arp and Konrad Rieck , title =. 2014. 2014 , url =. doi:10.1109/SP.2014.44 , timestamp =
-
[56]
Viterbi, A. , journal=. Error bounds for convolutional codes and an asymptotically optimum decoding algorithm , year=
-
[57]
Building and Interpreting Deep Similarity Models , journal =
Oliver Eberle and Jochen B. Building and Interpreting Deep Similarity Models , journal =. 2022 , doi =
work page 2022
-
[58]
Reduan Achtibat and Maximilian Dreyer and Ilona Eisenbraun and Sebastian Bosse and Thomas Wiegand and Wojciech Samek and Sebastian Lapuschkin , title =. CoRR , volume =
-
[59]
Xiaoqi Wang and Han. GNNInterpreter:. CoRR , volume =. 2022 , url =. doi:10.48550/arXiv.2209.07924 , eprinttype =. 2209.07924 , timestamp =
-
[60]
Huang and Nikhil Rao and Karthik Subbian and Shuiwang Ji , title =
Yaochen Xie and Sumeet Katariya and Xianfeng Tang and Edward W. Huang and Nikhil Rao and Karthik Subbian and Shuiwang Ji , title =. NeurIPS , year =
-
[61]
9th International Conference on Learning Representations,
Michael Sejr Schlichtkrull and Nicola De Cao and Ivan Titov , title =. 9th International Conference on Learning Representations,. 2021 , url =
work page 2021
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.