Pith. sign in

REVIEW 5 major objections 5 minor 21 references

RAW-Explainer: Post-hoc Explanations of Graph Neural Networks on Knowledge Graphs

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Connected subgraph explanations beat path-based KG rivals

desk verdict A plausible and coherent explainer for KG link prediction, but the robustness story needs an experiment on the explainer's own subgraphs before I'd trust it. read the letter →

arxiv 2506.12558 v1 pith:YAQ7KWWR submitted 2025-06-14 cs.LG stat.ML

Ryoji Kubo , Djellel Difallah This is my paper · ORCID
classification cs.LGstat.ML
keywords graphneuralnetworkexplainabilityknowledgecompletionlinkpredictionsubgraphexplanationsrandomwalkdistributionshiftrobustevaluatorfine-tuningevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

RAW-Explainer claims that the right way to explain a knowledge-graph link prediction is a small connected subgraph, and that existing explainers fail at this because they either ignore heterogeneous structure, do not guarantee connectivity, or evaluate subgraphs with a model that was never trained on anything that small. The paper's method learns a soft edge mask from entity and relation embeddings with a small neural network, then runs a personalized random walk to force the selected edges into one connected component around the query. To score candidate masks correctly, it trains a robust evaluator that drops edges with probability increasing with distance from the query head, so the evaluator stays accurate on subgraph-scale inputs. On FB15k-237 and WN18RR, fine-tuning the backbone on RAW-Explainer's explanations gives higher MRR than PaGE-Link and Power-Link and is competitive with PG-Explainer, while producing a single connected subgraph orders of magnitude faster.

What carries the argument

The central object is the pair ($\Phi_{\mathrm{dist}}^{eval}$, $\mathcal{L}_{PPR}$). $\Phi_{\mathrm{dist}}^{eval}$ is a GNN link predictor trained with distance-proportional edge dropping, so its predictions remain reliable on small subgraphs; it supplies the training signal for the mask. $\mathcal{L}_{PPR}$ converts the learned soft mask into a stochastic transition matrix with teleportation to query-relevant entities, runs a personalized random walk to convergence, then penalizes mask weight on edges whose endpoints are not both in the top-ranked entity set, enforcing that the final subgraph is connected around the query head and its likely tails.

What would settle it

Run a controlled study on synthetic KGs with planted ground-truth rule patterns. If RAW-Explainer's extracted subgraphs fail to cover the planted pattern at a high rate, or if the MRR of a backbone fine-tuned on its subgraphs does not exceed that of a random connected subgraph of the same budget, the method's central claim would be refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that explanation quality for KG link prediction hinges on three design choices working together: a heterogeneous, parameterized mask learner; a personalized random walk that enforces connectivity; and a distance-biased robust evaluator. The evaluator is trained by removing each edge with probability that grows with its hop distance from the query head, which mimics the small, local subgraphs that explanations actually are. The mask learner maximizes the evaluator's masked prediction while the random-walk loss rewards edges inside the query's reachable core and penalizes edges outside it, yielding a concise connected subgraph at any budget. The authors then assess explanations by downstream utility: they fine-tune the original backbone GNN on the extracted subgraphs and measure MRR, finding that RAW-Explainer outperforms the path-based KG baselines and matches a graph-wide explainer while guaranteeing connectivity and cutting inference time from hours to minutes.

Load-bearing premise

The whole method rests on the assumption that the distance-biased evaluator, trained on synthetic edge-dropped graphs, approximates the real distribution of explanatory subgraphs well enough that maximizing its masked prediction yields subgraphs the true backbone finds informative; the paper supports this only with the curves in Figure 3.

Editorial extensions

If this is right

  • Every KG link prediction can be explained by one connected subgraph at any user-chosen edge budget, making explanations readable as a small graph or path.
  • Replacing uniform edge dropping with distance-biased dropping improves any perturbation-based evaluator on heterogeneous graphs, not just this explainer.
  • The downstream fine-tuning protocol gives KG explainability a task-based metric, so methods can be compared without ground-truth explanations.
  • Parameterized mask learning makes collective explanation generation inductive and time-efficient: minutes instead of hours.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The evaluation protocol suggests a data-augmentation loop the authors did not run: fine-tuning the backbone on its own high-utility explanations could improve the link predictor itself, closing the explain-then-train circle.
  • Making the edge-drop probability depend on relation type as well as distance may sharpen the robust evaluator further, since informative edge patterns in KGs are relation-specific.
  • The connectivity guarantee is only as strong as the random walk's top-$\ell$ entity set; if an important edge lies outside that set, the loss actively penalizes it, so the method favors well-connected queries over scattered-but-important ones.
  • Re-running the PG-Explainer comparison with a connectivity penalty or post-hoc path extraction would isolate whether its higher MRR on FB15k-237 comes from genuinely better subgraphs or from exploiting disconnected latent structure.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. RAW-Explainer generates post-hoc subgraph explanations for knowledge graph link prediction. It parameterizes edge-mask learning with an MLP, regularizes masks through a Personalized PageRank random-walk objective to encourage connectivity, and replaces the backbone GNN with a distance-biased robust evaluator (Φdist_eval) when optimizing the mask. Explanations are evaluated by fine-tuning the backbone NBFNet on the extracted subgraphs and reporting MRR on test subgraphs, comparing against GNN-Explainer, PG-Explainer, PaGE-Link, and Power-Link on WN18RR and FB15k-237. The paper reports that RAW-Explainer outperforms KG-specific path baselines, is competitive with or behind PG-Explainer, and is substantially faster than instance-based methods.

Significance. The proposed downstream fine-tuning protocol (Section 4.5) is a practical way to compare explainers when ground truth is absent, and the parameterized, inductive mask learner is a sensible design for scalability. The efficiency gains are concrete and large (e.g., 17 minutes vs. 109–172 minutes on WN18RR). If the robustness of Φdist_eval to the explainer's own subgraph distribution were established, the method would offer a solid balance between efficiency, connectivity, and informativeness. However, the current evidence does not yet substantiate the surrogate-evaluator mechanism or the connectivity guarantee, and the reported point estimates are not accompanied by variance information. The contributions are promising but not fully established.

major comments (5)
  1. [§4.2, §5.2–5.3] The robust evaluator Φdist_eval is validated only on random edge-drop and ego-network perturbations, not on the distribution of subgraphs actually produced by the mask learner. Since the mask learner optimizes Eq. (5) against Φdist_eval, a systematic disagreement between Φdist_eval and the backbone Φ on those optimized subgraphs would undermine the factual-fidelity interpretation of the explanations, yet no experiment compares Φ and Φdist_eval on RAW-Explainer's output subgraphs. Please add a direct agreement or transfer experiment (e.g., correlation of masked predictions, or downstream MRR of masks trained with Φ vs. Φdist_eval) to support the claim that Φdist_eval 'generalizes to the subgraph distribution.'
  2. [Tables 3–4] All MRR results are reported as single point estimates with no standard deviations, confidence intervals, or significance tests. The evaluation protocol involves stochastic mask learning and GNN fine-tuning, and several margins are narrow (e.g., WN18RR, k=25: RAW-Explainer 0.482 vs. Power-Link 0.473). Without repeated runs and variance reporting, the comparative claims are not statistically assessable.
  3. [§5.1, §5.4] The manuscript does not report the values or tuning ranges of the free parameters introduced in the method, including α, the top-m predicted tails, top-l PPR entities, β_in/β_out, the Gumbel-Softmax temperature, the sparsity and entropy regularization weights, and the distance-decay rate for Φdist_eval. Section 5.4 states that hyperparameters are selected on the validation set, but without these details the results are not reproducible and the sensitivity of the method to key choices is unknown. A hyperparameter table and a sensitivity analysis for at least α and the β weights should be provided.
  4. [§4.3, §5.4] The paper claims that the random-walk objective 'ensures the connectivity of the output explanations,' and the FB15k-237 discussion states that RAW-Explainer 'returns a single connected component by design.' The described procedure—selecting top-l PPR nodes and then edges with both endpoints in that set—does not mathematically guarantee a connected induced subgraph, since two high-PPR nodes may be connected only through a node outside the top-l set. No connectivity metric is reported for the generated subgraphs (only PG-Explainer's average components is mentioned). Please prove the guarantee or report empirical connectivity statistics for all methods.
  5. [Introduction, Table 4] The introduction claims RAW-Explainer 'yields better explanatory subgraphs under limited edge budgets,' but Table 4 shows that PG-Explainer achieves the highest MRR at every budget on FB15k-237, with RAW-Explainer second. The discussion attributes PG-Explainer's advantage to disconnected subgraphs, but the claim as written is contradicted by the table. The claim should be qualified to 'better than path-based KG-specific baselines' or reworded to explicitly describe the trade-off between connectivity/interpretability and downstream MRR.
minor comments (5)
  1. [Eq. (7)] The teleportation matrix τ is described as binary but is combined with σ(Ω′) via a softmax; please clarify the normalization and the role of τ in the stochastic matrix construction.
  2. [Table 1] The entry 'RA W-Exp' appears with an awkward line break; it should read 'RAW-Exp' for consistency with the rest of the paper.
  3. [§5.4, Results Discussion] There is a typo, 'due to to the increased complexity,' which should be corrected to 'due to the increased complexity.'
  4. [Figure 3 caption] The caption says 'on different size of ego networks'; the grammar should be 'on different sizes of ego networks.'
  5. [§4.1–4.2] The notation for the evaluator is inconsistent: Section 4.1 uses Φeval, while Section 4.2 introduces Φdist_eval and Φunif_eval; please standardize the notation throughout.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the final explanation benchmark is external, and the self-referential evaluator loop is an empirical transfer risk, not a derivation that reduces to its inputs.

full rationale

RAW-Explainer's mask objective (Eq. 5) maximizes the masked prediction of the frozen robust evaluator Phi_dist_eval, and the PPR teleportation seeds in Section 4.3 are also taken from Phi_dist_eval's top-m predicted tails. This is a self-referential design, but it is not circular in the logical sense: the mask is not fitted to the downstream MRR, and the paper's final evaluation (Section 4.5) fine-tunes the backbone NBFNet on the extracted subgraphs and measures MRR on held-out test queries, which is an external benchmark that does not reduce to the training objective. Whether Phi_dist_eval accurately approximates the backbone on mask-optimized subgraphs is an empirical transfer assumption; the paper supports it with random edge-drop and ego-network experiments (Figures 2 and 3), and those experiments do not test the actual subgraph distribution the mask learner selects. That is a correctness and robustness concern, not a circularity: no equation in the paper is equivalent to its own input by construction. The only self-citation is reference [9], the authors' prior XGExplainer work, used to motivate the uniform edge-drop evaluator concept; the paper explicitly states it 'independently explores this claim' in Section 5.2, so that citation is not load-bearing. The central comparison with PaGE-Link, Power-Link, PG-Explainer, and GNN-Explainer is conducted under the same protocol and evaluator, and the reported improvements are empirical results rather than consequences of the definitions. Overall, no circular step meeting the required evidentiary standard was found.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central method rests on several unreported free parameters, a domain-specific choice of factual explanations, a connectivity-as-interpretability assumption, and the key premise that distance-biased edge dropping mimics the subgraph distribution. The latter is the most fragile; if it fails, the mask learner could exploit the evaluator rather than discover true explanatory structure.

free parameters (7)
  • teleportation weight alpha
    Controls blend of structural random walk vs teleportation in Eq.7; value not disclosed.
  • top-m predicted tails m
    Number of evaluator-predicted tails used to seed teleportation set rho in Section 4.3; not reported.
  • top-l PPR entities l
    Number of high-PPR entities defining E_in and E_out in Eq.8; not reported.
  • PPR loss weights beta_in, beta_out
    Balance reward for edges inside vs outside the PPR core in Eq.8; not reported.
  • Gumbel-Softmax temperature
    Temperature for approximating discrete edge sampling in Section 4.1; not reported.
  • sparsity and entropy regularization weights
    Regularizers pushing mask toward binary and small in Section 4.4; values not reported.
  • distance decay rate for Phi_dist_eval
    Rate at which edge-drop probability decays with distance from query head in Section 4.2; no formula or value given.
assumptions (6)
  • standard math Personalized PageRank power iteration converges to a query-dependent node distribution.
    Used in Section 4.3 to define top entities for connectivity loss; standard convergence result assumed.
  • standard math Gumbel-Softmax provides a differentiable approximation to discrete edge sampling.
    Used in Section 4.1 to obtain soft mask in [0,1].
  • domain assumption Factual (sufficiency) explanations are the right criterion for KG link prediction.
    Section 3 opts for phenomenon-focus vs model-focus per Amara et al. [2]; this choice is not derived.
  • domain assumption Connected subgraphs are more interpretable than disconnected ones.
    Stated in Sections 1 and 4.3; no human study or quantitative interpretability metric provided.
  • ad hoc to paper Training an evaluator with distance-biased edge dropping approximates the distribution of explanatory subgraphs.
    Core premise of the proposed robust evaluator in Section 4.2; supported only by the empirical curves in Figure 3, no theoretical justification.
  • domain assumption Fine-tuning the backbone GNN on explanation subgraphs and measuring MRR measures explanation quality.
    Proposed evaluation protocol in Section 4.5; the paper argues intuition but does not validate it against ground-truth explanations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RAW-Explainer: Post-hoc Explanations of Graph Neural Networks on Knowledge Graphs." pith.science (2026). https://pith.science/paper/YAQ7KWWR

@misc{pith2026250612558,
  author       = {Pith},
  title        = {Pith review of: RAW-Explainer: Post-hoc Explanations of Graph Neural Networks on Knowledge Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YAQ7KWWR}},
  note         = {Machine review of arXiv:2506.12558}
}
read the original abstract

Graph neural networks have demonstrated state-of-the-art performance on knowledge graph tasks such as link prediction. However, interpreting GNN predictions remains a challenging open problem. While many GNN explainability methods have been proposed for node or graph-level tasks, approaches for generating explanations for link predictions in heterogeneous settings are limited. In this paper, we propose RAW-Explainer, a novel framework designed to generate connected, concise, and thus interpretable subgraph explanations for link prediction. Our method leverages the heterogeneous information in knowledge graphs to identify connected subgraphs that serve as patterns of factual explanation via a random walk objective. Unlike existing methods tailored to knowledge graphs, our approach employs a neural network to parameterize the explanation generation process, which significantly speeds up the production of collective explanations. Furthermore, RAW-Explainer is designed to overcome the distribution shift issue when evaluating the quality of an explanatory subgraph which is orders of magnitude smaller than the full graph, by proposing a robust evaluator that generalizes to the subgraph distribution. Extensive quantitative results on real-world knowledge graph datasets demonstrate that our approach strikes a balance between explanation quality and computational efficiency.

Figures

Figures reproduced from arXiv: 2506.12558 by the authors.

Figure 1
Figure 1. End-to-end evaluation protocol for GNN explanations in knowledge graph completion. Explanations are assessed by [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Performance evaluation of Φ, Φ𝑒𝑣𝑎𝑙 on different probability of random edge-drop [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Performance evaluation of Φ, Φ 𝑢𝑛𝑖 𝑓 𝑒𝑣𝑎𝑙 , and Φ 𝑑𝑖𝑠𝑡 𝑒𝑣𝑎𝑙 on different size of ego networks around the head entity. 5.3 Comparing Variants of Model Evaluators on Distance In this experiment, we compare three GNN evaluators: the backbone model Φ, an evaluation model Φ unif 𝑒𝑣𝑎𝑙 trained using a uniform edge￾drop distribution, and a distance-biased evaluator Φ dist 𝑒𝑣𝑎𝑙 . Unlike Φ unif 𝑒𝑣𝑎𝑙 , which removes edges unif… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 19 canonical work pages

  1. [1]

    Scientific Data 10(1), 144 (2023)

    Agarwal, C., Queen, O., Lakkaraju, H., Zitnik, M.: Evaluating explainability for graph neural networks. Scientific Data 10(1), 144 (2023)

    work page 2023

  2. [2]

    arXiv preprint arXiv:2206.09677 (2022)

    Amara, K., Ying, R., Zhang, Z., Han, Z., Shan, Y., Brandes, U., Schemm, S., Zhang, C.: Graphframex: Towards systematic evaluation of explainability methods for graph neural networks. arXiv preprint arXiv:2206.09677 (2022)

    arXiv 2022

  3. [3]

    Baltatzis, V., Costabello, L.: Kgex: Explaining knowledge graph embeddings via subgraph sampling and knowledge distillation. In: LoG. Proceedings of Machine Learning Research, vol. 231, p. 27. PMLR (2023) Ryoji Kubo and Djellel Difallah

    work page 2023

  4. [4]

    In: NIPS

    Bordes, A., Usunier, N., García-Durán, A., Weston, J., Yakhnenko, O.: Translating embeddings for modeling multi-relational data. In: NIPS. pp. 2787–2795 (2013)

    work page 2013

  5. [5]

    Chang, H., Ye, J., Lopez-Avila, A., Du, J., Li, J.: Path-based explanation for knowl- edge graph completion. In: KDD. pp. 231–242. ACM (2024)

    work page 2024

  6. [6]

    In: AAAI

    Dettmers, T., Minervini, P., Stenetorp, P., Riedel, S.: Convolutional 2d knowledge graph embeddings. In: AAAI. pp. 1811–1818. AAAI Press (2018)

    work page 2018

  7. [7]

    In: AISTATS

    Jethani, N., Sudarshan, M., Aphinyanaphongs, Y., Ranganath, R.: Have we learned to explain?: How interpretability methods can learn to encode predictions in their interpretations. In: AISTATS. Proceedings of Machine Learning Research, vol. 130, pp. 1459–1467. PMLR (2021)

    work page 2021

  8. [8]

    IEEE Data Eng

    Kakkad, J., Jannu, J., Sharma, K., Aggarwal, C., Medya, S.: A survey on explain- ability of graph neural networks. IEEE Data Eng. Bull. 46(2), 35–63 (2023)

    work page 2023

Show all 21 references
  1. [9]

    In: Proceedings of the 2024 SIAM International Confer- ence on Data Mining (SDM)

    Kubo, R., Difallah, D.: Xgexplainer: Robust evaluation-based explanation for graph neural networks. In: Proceedings of the 2024 SIAM International Confer- ence on Data Mining (SDM). pp. 64–72. SIAM (2024)

    2024

  2. [10]

    In: NeurIPS (2020)

    Luo, D., Cheng, W., Xu, D., Yu, W., Zong, B., Chen, H., Zhang, X.: Parameterized explainer for graph neural network. In: NeurIPS (2020)

    2020

  3. [11]

    Technical Report 1999-66, Stanford InfoLab (November 1999)

    Page, L., Brin, S., Motwani, R., Winograd, T.: The pagerank citation ranking: Bringing order to the web. Technical Report 1999-66, Stanford InfoLab (November 1999)

    1999

  4. [12]

    In: ESWC

    Schlichtkrull, M.S., Kipf, T.N., Bloem, P., van den Berg, R., Titov, I., Welling, M.: Modeling relational data with graph convolutional networks. In: ESWC. Lecture Notes in Computer Science, vol. 10843, pp. 593–607. Springer (2018)

    2018

  5. [13]

    In: CVSC

    Toutanova, K., Chen, D.: Observed versus latent features for knowledge base and text inference. In: CVSC. pp. 57–66. ACL (2015)

    2015

  6. [14]

    In: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020

    Vashishth, S., Sanyal, S., Nitin, V., Talukdar, P.P.: Composition-based multi- relational graph convolutional networks. In: 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net (2020), https://openreview....

    2020

  7. [15]

    In: NeurIPS

    Ying, Z., Bourgeois, D., You, J., Zitnik, M., Leskovec, J.: Gnnexplainer: Generating explanations for graph neural networks. In: NeurIPS. pp. 9240–9251 (2019)

    2019

  8. [16]

    IEEE Trans

    Yuan, H., Yu, H., Gui, S., Ji, S.: Explainability in graph neural networks: A taxo- nomic survey. IEEE Trans. Pattern Anal. Mach. Intell. 45(5), 5782–5799 (2023). https://doi.org/10.1109/TPAMI.2022.3204236

    2023

  9. [17]

    In: ICML

    Yuan, H., Yu, H., Wang, J., Li, K., Ji, S.: On explainability of graph neural networks via subgraph explorations. In: ICML. Proceedings of Machine Learning Research, vol. 139, pp. 12241–12252. PMLR (2021)

    2021

  10. [18]

    Zhang, S., Zhang, J., Song, X., Adeshina, S., Zheng, D., Faloutsos, C., Sun, Y.: Page-link: Path-based graph neural network explanation for heterogeneous link prediction. In: WWW. pp. 3784–3793. ACM (2023)

    2023

  11. [19]

    Zhang, Y., Yao, Q.: Knowledge graph reasoning with relational digraph. In: WWW. pp. 912–924. ACM (2022)

    2022

  12. [20]

    In: NeurIPS (2023)

    Zhu, Z., Yuan, X., Galkin, M., Xhonneux, L.A.C., Zhang, M., Gazeau, M., Tang, J.: A*net: A scalable path-based reasoning approach for knowledge graphs. In: NeurIPS (2023)

    2023

  13. [21]

    In: NeurIPS

    Zhu, Z., Zhang, Z., Xhonneux, L.A.C., Tang, J.: Neural bellman-ford networks: A general graph neural network framework for link prediction. In: NeurIPS. pp. 29476–29490 (2021)

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.