acceptodds
Under review as a conference paper at ICLR 2027

Statistical Subgraph Explanations for Graph Neural Networks via Latent Node Inference

Abstract

Post-hoc explanations of graph neural networks (GNNs) often rely on learned edge masks or combinatorial search, leaving the statistical evidence for the selected structure implicit. We recast explanatory subgraph discovery as latent node inference and investigate when departures from background connectivity support explanations of a fixed GNN. We develop two complementary methods without training an auxiliary neural explainer. HierExplainer uses variational inference under a Bernoulli block model for each specific target to estimate posterior node-membership probabilities. HypoExplainer screens prediction label pairs of GNN using edge distribution after Bonferroni correction tests against a supervised background reference, then constructs explanations within each target’s computation graph. Our theory distinguishes structural recovery from predictive faithfulness: under explicit compatibility assumptions, HierExplainer’s posterior objective combines prediction fidelity with structural regularization, while HypoExplainer admits an asymptotic recovery guarantee under edge independence and minimal signal conditions. Experiments assess recovery, faithfulness, stability, and resource cost on synthetic motif benchmarks and a protein-protein interaction network.

Primary area:
interpretability and explainable AI
Keywords:
GNN Explainability, Post-hoc Explanation, Subgraph Explanation, Statistical Inference, Hypothesis Testing, Stochastic Block Model

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_JjK1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HierExplainer and HypoExplainer for post-hoc GNN explanation that formulate explanatory subgraph discovery as latent-node inference. HierExplainer uses variational inference under a generative model, while HypoExplainer uses hypothesis testing to identify structures whose connectivity differs significantly from the background.

Critical strengths
  1. The paper formulates subgraph explanation as a statistical problem, which an interesting alternative approach to commonly used XAI methods.

  2. The proposed methods are well motivated and provide a potential valuable statistical way for further XAI research for GNN.

Critical weaknesses
  1. The method critically relies on the assumptions, particularly the very strong assumptions 2, 4 and 6, but their validity on real-world graphs is not sufficiently examined. I'm concerned about the applicability of the method and its theoretical guarantees to real-world data.

  2. The paper has no clear definition of the Saliency baseline and doesn't specify how the other baseline methods are implemented. Besides, for the GIN experiments, different methods use different target cohorts and F1 definitions (Line 789), which raises concern about whether the F1 values are comparable.

Questions for authors
  1. How are the sizes of the subgraph determined for each baselines in Fig 5? For example, what is the selection threshold for GNNExplainer's edge scores? This is important for the claim that baselines' subgraph is less compact than the proposed method (line 358)

  2. How do the proposed methods compare with other popular post-hoc GNN explanation methods, such as Gradient × Input, Integrated Gradients, GNN-LRP, etc.?

  3. Figure 4 is too small.

Reviewer_7ATKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes two statistical approaches to explaining GNN node predictions. HierExplainer uses probabilistic modeling and variational inference, while HypoExplainer identifies candidate explanatory structures through hypothesis testing against background connectivity. Experiments evaluate explanation quality, stability, and computational cost on synthetic graphs and a protein–protein interaction network.

Critical strengths
  1. The paper offers a statistical perspective on identifying explanatory subgraphs without training an additional neural explainer.
  2. The experiments examine several aspects of explanation quality using both synthetic and real-world data.
Critical weaknesses
  1. The research motivation is insufficiently clear. The paper does not adequately explain the specific limitations of existing methods that motivate the proposed approach.
  2. The novelty is not sufficiently established. The related-work discussion only covers GNN explanation methods published up to 2022, and the experimental baselines are relatively old. Comparisons with more recent relevant approaches are missing.
  3. The manuscript lacks sufficient organization and proofreading, which affects the clarity of its technical presentation. Figure 5 is particularly poorly presented: the dataset corresponding to each row is not clearly labeled, and the layout makes comparisons difficult. The visualization also does not clearly demonstrate the advantages of HypoExplainer.
  4. The interpretation of Figure 6 is insufficient. It remains unclear what the dense, clique-like structure reveals and why it should be considered a better explanation than the outputs of other explainers.
Questions for authors
  1. What exactly does “background connectivity” mean, and how is it determined?
  2. What distinguishes the proposed methods from more recent GNN explanation approaches? Why were these approaches not included in the comparisons?
Reviewer_JiLuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes two statistical approaches for post-hoc GNN explanation. HierExpaliner uses variational inference to identify explanatory graphs, while HypoExpaliners uses hypothesis testing to find statistically significant patterns. The evaluations have been done on syntehtic and real-world graphs.

Critical strengths

It explores a different perspective on GNN explanation by explicitly modeling graph connectivity and comparing it with background structures.

Critical weaknesses
  1. The motivation and novelty are not very convincing. The paper argue that existing GNN explainers provide limited statistical evidence about whether selected structures differ from background connectivity. However, probabilistic modeling are not new to GNN explanation. For example, PGExplainer already models distributions over explanatory subgraphs using Bernoulli variables. Although this paper introduces different statistical assumptions, I'm not sure what fundamental problem it addresses that existing methods cannot solve. I think the authors should better explain what makes their approach different and why these statistical methods are necessary.

  2. The connection between statistical significance and GNN explanation is not convincing enough. The proposed methods identify subgraphs based on their statistical connectivity patterns. However, just because a subgraph is statistically different from the background doesn't necessarily mean the GNN actually relies on it to make predictions. Unlike methods such as PGExplainer, which directly optimize prediction fidelity, the proposed methods mainly focus on structural inference. The authors try to establish this connection through their theoretical results, but these results rely on strong assumptions. I'm not sure how realistic these assumptions are in practice, and I don't think the experiments provide enough evidence to show that recovering statistically distinctive structures consistently leads to faithful GNN explanations.

  3. The experimental evaluation is not comprehensive enough. HierExplainer is mainly evaluated on synthetic SBM-generated graphs, which already follow its modeling assumptions. Its experiments focus more on structural recovery than on showing whether it can faithfully explain GNN predictions compared with existing explainers. Although HypoExplainer is evalauted more extensively, its performance varies quite a bit across different graph structures and background connectivity settings, which makes me question how well it can generalize.

Questions for authors

see the weakness above.

Reviewer_saxgOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper formulates post-hoc GNN subgraph explanation as latent node inference. It introduces HierExplainer, based on variational inference under a Bernoulli block model, and HypoExplainer, based on connectivity tests against a supervised background reference. Theoretical results connect structural recovery to predictive faithfulness under explicit assumptions. Experiments examine explanation recovery, faithfulness, stability, and computational cost on synthetic and real graphs.

Critical strengths

I think the approach in this paper has the potential to address some inherent limitations of GNNs, and its theoretical framework aligns well with modeling real-world scenarios.

Critical weaknesses
  1. The connection between structural recovery and model faithfulness relies on strong assumptions.
  2. The validity of the statistical tests under graph dependence remains insufficiently established.
  3. Differences in evaluation protocols limit comparisons with baselines.
Questions for authors
  1. Could you test whether the explanations change when a model relies on node features rather than graph structure?
  2. Reporting false-positive rates across repeated graph samples with dependent edges and varying node degrees would help clarify the paper’s contribution.
  3. Could you compare the methods on the same target nodes, using consistent evaluation metrics and matched explanation sizes? The current evaluation setup is a little hard to follow.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.