acceptodds
Under review as a conference paper at ICLR 2027

Graph Additive Models disentangling feature, topology, and interactions

Abstract

Graph Neural Additive Networks (GNANs) have emerged as an effective, interpretable-by-design graph architecture for learning on graphs. They achieve model transparency by combining per-feature shape functions with a simple distance-based reweighting. Despite the effectiveness of GNAN, we identify the following limitations: i) node-wise relevance scores are computed as a weighted aggregation over all graph nodes, allowing irrelevant nodes to inherit spurious relevance from relevant ones; ii) graph topology is encoded only through shortest-path distances, and structurally distinct graphs may have identical distances; iii) feature and distance information are combined multiplicatively, thus entangling feature, distance, and interaction effects into a single, rescalable term. To address these limitations, we introduce AMIGA, which represents topology through a vocabulary of mined motifs and decomposes each node’s contribution into three additive channels: a feature channel, a structural channel, and a feature–structure interaction channel. We center these channels around a training-data baseline that gives every graph equal weight, regardless of its size, making the decomposition well defined and unique. Across ten graph classification and regression benchmarks, AMIGA outperforms GNAN on most datasets and narrows the gap to standard black-box GNNs. Because every channel is centered and attached to a single node, AMIGA’s explanations stay local to the node that generates them and reconstruct the model’s output exactly while providing interpretability at different levels of granularity. Code is publicly available online.

Primary area:
interpretability and explainable AI
Keywords:
self-explainable models, graph neural networks, interpretability, generalized additive models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rHYPOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces AMIGA, an extension of the self-interpretable model GNAN, which discovers structural motifs as a vocabulary and decomposes contributions into feature, structural, and feature-structure interaction channels.

Experiments on graph classification/regression datasets show that the proposed method outperforms its predecessor GNAN and achieves performance closer to standard black-box GNNs.

Critical strengths
  1. Using a motif vocabulary to capture structural signals is a sound and natural design.
  2. This paper identifies several limitations of the existing model GNAN and presents a non-trivial extension.
Critical weaknesses
  1. Considering the number of motifs, features, and motif-feature interactions, the resulting contributions may still be difficult to analyze, even if each component is individually interpretable. This limits the practical interpretability of the proposed method.

  2. As the proposed method does not generally outperform standard black-box GNNs, further validation of its explanation capability beyond BA-based synthetic datasets would help strengthen the significance of this work. For example, the GNNExplainer paper[1] also incorporates tree-based graphs with ground-truth explanations.

  3. The complexity of the motif-based approach is concerning for larger motifs or graphs. Currently, the paper only incorporates graphlets of up to five nodes, potentially limiting its ability to capture more complex structural patterns.

[1]: GNNExplainer: Generating explanations for graph neural networks

Questions for authors
  1. Could the authors briefly discuss the computational complexity of the proposed method, particularly motif discovery, with respect to graph size and the number/size of graphlets?
5 comments
AuthorsOpenReview ↗

Response to Reviewer rHYP: Practical Interpretability, Additional Datasets, and Motif Scalability

Thank you for your detailed review and your valuable time. Below, we address your concerns.

W1: Practical Interpretability

Thank you for raising this concern. We agree that inspecting all feature–motif contributions as a flat list could be difficult. However, AMIGA explanations can be better interpreted hierarchically, where users first compare the graph-level feature, structural, and interaction contributions (ΦF,ΦS,ΦI)(\Phi^F,\Phi^S,\Phi^I), and then rank them by magnitude and inspect only the leading terms within the dominant channel. We provide two concrete examples of this below:

Example 1: Consider a trained AMIGA on the TREE–GRID dataset. For this example, the prediction is entirely driven by the structural channel (ΦS=6.84\Phi^S=6.84 ), while the feature and interaction channels make no contribution (ΦF=0\Phi^F=0 and ΦI=0\Phi^I=0 ). Therefore, the explanation can focus exclusively on the graph structure.

Example 2: Consider a trained AMIGA on the MUTAGENICITY dataset. For a graph predicted as non-mutagenic, the structural channel provides the largest contribution to the prediction, accounting for 93.893.8% of the total. In contrast, for a graph predicted as mutagenic, the interaction channel dominates, accounting for 95.995.9% of the total contribution. This illustrates how AMIGA decomposes individual predictions and identifies which aspects of the graph drive each outcome.

The number of inspectable elements is further limited by:

  • keeping only statistically significant graphlets
  • treating each one-hot encoded feature as a single categorical variable. For example, AMIGA treats the 14 coordinates encoding atom type in Mutagenicity as one variable, while GNAN and G-NAMRFF train 14 different shape functions;
  • expressing each feature–motif interaction as the product of a feature response and a motif response, so that every pair remains individually inspectable without learning a separate function for each pair
  • using a sparsity penalty to push uninformative effects toward zero.

We added this remark and the additional experiments in the revised manuscript, in Section 4.4 and Appendix C.2.


W2: Additional datasets

We thank the reviewer for the comment. Following your suggestion, we included two additional tree-based graph-classification experiments: tree versus tree + cycle, and tree versus tree + grid. Each dataset contains 1,0001,000 balanced examples, with comparable node-count distributions and constant features. The planted motif nodes provide, therefore, ground-truth explanations and evaluate explanation AUROC on all positive test graphs for five seeds.

ModelTree + CycleTree + Grid
GIN + GNNExplainer0.493 ± 0.2270.608 ± 0.109
GIN + IG0.636 ± 0.4770.698 ± 0.280
GNAN0.653 ± 0.2300.737 ± 0.193
G-NAMRFF0.379 ± 0.1880.423 ± 0.140
AMIGA0.668 ± 0.0420.997 ± 0.001

Under the evaluated configurations, AMIGA and GIN both achieve near-perfect predictive AUROC, but their explanation AUROC differs, and AMIGA achieves the best performance in both datasets. We added these additional experiments in Appendix C.1 of the revised manuscript.


W3/q1: Computational Complexity and Motif Scalability.

Thank you for raising this concern. We have expanded the complexity discussion and added the following results to Appendix B.2:

Computational cost. To enumerate graphlets, we use RAND-ESU [1]. For fixed graphlet size qq and bounded degree, the complexity scales linearly with the number of nodes [2]. We report runtimes on ER graphs with increasing ∣V∣|V| for q∈{3,4,5,6}q \in \{3,4,5,6\} in Appendix B.2. For a fixed graph with 10,00010,000 nodes, the results show a runtime ranging from 0.0080.008 for q=3q=3 up to 1.81.8 seconds for q=6q=6, with an approximately linear trend.

On our largest benchmark (QM9), enumerating all 2929 candidates over the 91,58191,581 training graphs takes 2.522.52 seconds. Note that this cost is for the motif discovery part, which is paid once at pre-processing time.

Motif size. Although AMIGA is not restricted to a particular motif size and can use any motif vocabulary supplied by the counting backend, we choose qmax=5q_{max}=5 for our experiments. This choice is mainly motivated by the following:

  • Interpretability: Small graphlets capture informative local structures while preserving reliable statistical support [3] and human-readable structural representations [4].
  • Number of graphlets: The number of connected graphlets grows rapidly, from 2 at q=3q=3 to 6 at q=4q=4, 21 at q=5q=5, and 112 at q=6q=6.
  • Meaningful support: Bigger graphlets occur more sparsely across graphs, which weakens their empirical support [5].

Thank you again; your remark helped in making our paper more rigorous! Please let us know if our answers address your concerns, or whether we can provide further clarifications.

AuthorsOpenReview ↗

References:

[1] Efficient detection of network motifs. IEEE/ACM transactions on computational biology and bioinformatics. 2006.

[2] Efficient graphlet kernels for large graph comparison. Artificial intelligence and statistics. 2009.

[3] Graft: An efficient graphlet counting method for large graph analysis. IEEE Transactions on Knowledge and Data Engineering. 2014.

[4] Scalability of network visualisation from a cognitive load perspective. IEEE transactions on visualization and computer graphics. 2020

[5] Pinar Yanardag and S. V. N. Vishwanathan. The Structurally Smoothed Graphlet Kernel, 2014.

I appreciate the clarifications and the additional effort the authors put into the additional datasets. My concerns (mainly W1 and the explanation ability of the proposed method) have been addressed.

However, I remain concerned about the heavy limitations of using graphlet motifs. For example, the computational complexity is linear if the maximum node degree is bounded, while this assumption does not hold for many types of graphs, such as BA graphs. Furthermore, motifs require sufficient statistical support to be meaningful, which limits their applicability to small, local structural patterns.

Overall, I have raised my rating to 2.

Reviewer_RduPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AMIGA is an interpretable graph predictor built from node attributes and mined motif participation counts. It separates each node score into feature, structural, and factorized interaction channels, centered under a graph-balanced product reference. These channels reconstruct the prediction exactly. Evaluation includes ten prediction benchmarks, synthetic effect recovery, and explanation recovery on two planted-motif datasets.

Critical strengths
  • The explanations expose the actual predictive computation. The explicit centering convention separates the three channels, and the paper carefully states the scope of its uniqueness guarantee. This provides a clear, auditable alternative to explanations based on a surrogate model.
  • The evaluation tests both prediction and explanation recovery. AMIGA improves on GNAN on eight of ten benchmarks and G-NAMRFF on nine, while recovering known effects and planted explanatory nodes. This supports its usefulness beyond the decomposition identities alone.
Critical weaknesses
  • The interaction channel is restricted to a single separable product, A(x)B(c). This ties all feature-motif interactions to shared factors and cannot represent several independent interaction patterns in general. The effect-recovery experiment contains only one active interaction, so it does not test this limitation or explain the performance gaps on tasks such as PROTEINS and QM9-alpha.
  • The uniqueness theorem applies to a fixed function on the product support, which can contain unobserved feature-topology combinations. It does not establish that equally predictive fits learn stable channel attributions when features and topology are correlated. Since the recovery experiment makes them independent, this important interpretation setting remains untested.
Questions for authors

Please see the weaknesses above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer RduP: Interaction Rank and Feature–Topology Dependence

Thank you for your detailed review and your valuable time. Below, we address your concerns.

W1: Factorized Interaction Channel.
We agree that a higher number of separable products of the interaction channel is expected to be more expressive. To test this empirically, we evaluate our model by increasing this number, hence by computing I(x,c)=∑r=1RAr(x)Br(c)I(x,c)=\sum_{r=1}^{R}A_r(x)B_r(c) with R=2,4R=2,4. We will refer to RR as the rank of the interaction channel. Results are described below:

Recovering-known-effects experiment. We extend the synthetic task of Section 5.1 with a second independent interaction and compare ranks. A rank equal to 22 reduces the test MAE from 1.787 to 1.290 (28%), while higher ranks provide little benefit on the original single-interaction task.

Real-world experiment. On the benchmarks of Section 5.3, with hyperparameters tuned separately for each rank, higher ranks improve mean performance on several tasks, including PROTEINS (AUROC: 0.733→0.7570.733 \rightarrow 0.757) and QM9$-\alpha$ (MAE: 3.309→2.9503.309 \rightarrow 2.950).

Overall, higher ranks preserve exact additive explanations but increase the number of functions to inspect: from d+Md+M to R(d+M)R(d+M) for the interaction channel. This makes explanations substantially harder to read. We therefore retain rank one as a deliberate default, trading some expressivity for interpretability. Full settings and results are reported in Appendix C.3.


W2: Feature–Topology Dependence. Thank you for highlighting this distinction. We agree that the theorem establishes uniqueness for a fixed function on the product support, but does not guarantee recovery from correlated data. This aspect was not properly discussed in the original version, and we now state this explicitly in Section 5.1.

To showcase the practical consequences of this, we performed the following additional experiment: we repeat the synthetic recovery task of Section 5.1 while varying the feature–topology correlation, using five correlation levels and random seeds. The results show that recovery remains accurate under weak correlation. Under strong correlation, instead, the recovery worsens whilst maintaining competitive predictive performance. This worsening in performance is expected, as additive models implicitly assume the independence among feature-feature and topology-feature interactions. Full results and details are reported in Appendix C.5.

Thank you again; your remark helped in making our paper more rigorous.

AuthorsOpenReview ↗

Response to Reviewer: asking clarifications

Dear reviewer,

Thank you once again for your feedback. As the discussion period is about to reach its end, we kindly ask whether you have any remaining comments or whether our answers have resolved your concerns.

Best,

Thank you for the detailed response and additional experiments. The new analyses on interaction rank and feature–topology correlation adequately address my main concerns and help clarify the limitations of the proposed method. I appreciate these improvements and will maintain my original assessment.

Reviewer_kjMsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AMIGA is a graph additive model that separates predictions into feature effects, strutural effects based on motif counts, and feature–motif interactions. The model uses a training-derived motif vocabulary and centers its shape functions under a graph-balanced reference distribution. These components can then be inspected at different levels and summed to recover the model score. The experiments cover synthetic effect recovery, explanation quality on two synthetic datasets, and prediction on ten benchmarks. The reported results generally improve on existing graph additive models, although black-box GNNs remain stronger on most real-world tasks.

Critical strengths
  • I like the separation of feature, structure, and interaction effects. It gives a concrete way to inspect what the fitted model is doing, and the contributions sum exactly to its score. The uniqueness result is also correct under the stated product-reference assumptions. This is a usefull extension of additive modeling to graph data.

  • The experiments examine more than prediction accuracy. The recovery example and channel ablations help explain how the model works. In particular, the different roles of the structural and interaction channels on BA-Reg 2 and BA-Reg 1 support the motivation for including both.

Critical weaknesses
  • I am not convinced by the claim in Section 4.3 that centering only removes an ambiguity in the representation. A constant g0g_0 in the node score becomes ∣V(G)∣g0|V(G)|g_0 after pooling. When graphs have different sizes, this cannot generally be absorbed into a single graph intercept. Equation (3) and the implementation do not include this size-dependent term. A simple example is C6C_6 and C7C_7 with identical constant node features. Every node has the same participation counts for all connected graphlets of sizes 3–5. All centered shapes therefore vanish, so the model gives both graphs the same prediction and cannot represent their node counts. This does not invalidate Theorem 1, but it means centering restricts the model class. Appendix B.5 already acknowledges the fixed-size requirement when absorbing a constant. The main formulation needs the same clarification.

  • Some reported results do not match the supplied experiment records. For PTC-MR, the five saved runs give mean AUROC 0.57436 and population standard deviation 0.05688, while Table 2 gives 0.574±0.0260.574 \pm 0.026. For QM9-α, the saved runs give 3.30883±0.277953.30883 \pm 0.27795, compared with 3.35±0.033.35 \pm 0.03. Using sample standard deviations does not explain these differences. The saved search space also includes dropout 0.3, which is absent from the search space described in Appendix B.3. These may be version mismatches, but the discrepencies need to be resolved before the reported comparisons can be assessed reliably.

Questions for authors
  • Is excluding the ∣V(G)∣g0|V(G)|g_0 term an intentional modeling choice? If so, please explain the restriction it introduces. Would including a learnable graph-size term change the results on variable-size datasets?

  • Could you identify the runs that produced Tables 2 and 6 and reconcile their metrics and training settings with the supplied records?

  • The original GNAN paper already gives a source-node attribution in Section 3, Equation (2). Summing it over features avoids the specific blurring shown in Figure 1a and still reconstructs the graph score exactly. What happens to the Table 1 comparison when this attribuition is used?

3 comments
AuthorsOpenReview ↗

Response to Reviewer kjMs: Graph-Size Dependence, Experimental Consistency, and GNAN Attribution

Thank you for your detailed review and your valuable time. Below, we address your concerns.

W1/Q1: Centering and Graph-Size Dependence.

Thank you for pointing this out. We agree: omitting g0∣V(G)∣g_0|V(G)| restricts the model class when graph sizes vary. Our original formulation uses only the three centered node-level channels, but we did not explain this restriction. We have clarified this in Section 4.3 of the revised manuscript.

Global graph descriptors – like ∣V(G)∣g0|V(G)|g_0 that you propose – can be included through an additional additive channel. In a new appendix C.4, we evaluate the example you suggested: a learnable node-count term. This fixes the limitation above and resolves the C6/C7C_6/C_7 counterexample while preserving an exact additive explanation.
Using five seeds, the original splits, and motifs, this extension improves mean performance on seven out of eight tasks. Thus, including graph size can improve performance at the cost of an additional channel, which can make explanations harder to analyse. The benefits of this extension, therefore, depend on the task. We added this additional discussion in Section 4.3 and in Appendix C.4.

We thank the reviewer for this suggestion, which improved the discussion in our manuscript.


W2/Q2: Reported Metrics and Experimental Settings
Thank you for identifying these discrepancies. We have corrected Tables 2 and 6 and revised Appendix B.3 to reflect the documented training settings, search spaces, and model-selection procedures.


Q3: GNAN comparison.
It is true that Eq. 2 gives a source-node attribution that fixes the limitation in Figure 1a, but what we note is that GNAN and G-NAMRFF explicitly use receiver-based attribution – i.e., aggregating message contributions at the receiving nodes – to extract node-wise relevance scores; see Section 3 of both GNAN and G-NAMRFF. Our contribution clarifies that this is suboptimal, and we believe it is a remark worth of notice, as it directly informs practitioners on which aggregation schema to use for future applications of graph additive models.

Following your suggestion, we ran the additional experiments using Eq. 2 of GNAN and reported the results below and in Table 1 of the revised manuscript. The results show that AMIGA still outperforms GNAN.

Re-evaluating the saved checkpoints gives:

Model / attributionBA-Reg 1BA-Reg 2
GNAN–Receiver (Eq. 3 of [1])0.71 ± 0.120.45 ± 0.16
GNAN–Source (Eq. 2 of [1])0.80 ± 0.110.54 ± 0.15
AMIGA0.99 ± 0.021.00 ± 0.00

[1] The intelligible and effective graph neural additive network. NeurIPS 2024

AuthorsOpenReview ↗

Response to Reviewer: asking clarifications

Dear reviewer,

Thank you once again for your feedback. As the discussion period is about to reach its end, we kindly ask whether you have any remaining comments or whether our answers have resolved your concerns.

Best,

Thank you for the detailed response and the additional experiments.

The revision to Section 4.3 addresses my concern about centering: it now explicitly states that omitting the constant node contribution restricts expressivity when graph sizes vary. The additional node-count channel in Appendix C.4 provides a direct way to address this limitation while keeping the decomposition additive. I should also clarify my original counterexample: the conclusion that all centered shapes vanish assumes a training reference consisting of the constant-attribute C6/C7C_6/C_7 graphs. Identical node inputs in these two graphs alone would not imply zero contributions under an arbitrary training reference. This condition should have been explicit in my review.

The corrected PTC-MR and QM9-α entries, together with the updated training and search settings in Appendix B.3, address the specific discrepancies I raised between the manuscript and the supplied records.

Thank you also for adding the GNAN source-attribution comparison. My concern was that source attribution was already available in GNAN and should be considered when discussing the blurring issue. The revised discussion distinguishes the two attribution conventions more clearly, and the additional results show that AMIGA retains its explanation-AUROC advantage on both BA-Reg benchmarks.

These revisions substantially address the concerns in my original review and make my assessment more positive. I have no further questions on these points and will take the changes into account when updating my overall recommendation: rating from 2 to 3.

Comments on the paper

2 comments

Thank you to the reviewer who has already followed up on the authors' responses. For reviewers who have not yet done so, please post a brief follow-up acknowledging the authors' responses within the next 24 hours, so the authors have time to respond before discussion closes. In your reply, please indicate which main concerns have been addressed, what remains, and whether your overall assessment changes or stays the same.

AuthorsOpenReview ↗

Summary of reviewers’ feedback, revisions, and final responses

We thank the reviewers for their careful and constructive feedback. We also appreciate that they recognized the following strengths:

  • AMIGA provides clear and auditable explanations with an exact decomposition of its predictions (kjMs, RduP).
  • The motif vocabulary is a sound and natural design (rHYP), and AMIGA is a useful, non-trivial extension of GNAN (kjMs).
  • Our evaluation supports the usefulness of the method (RduP).

In response to the reviews, we:

  • corrected Tables 2 and 6 and aligned the experimental settings with the provided records (kjMs);
  • clarified the graph-size limitation, added an optional graph-size channel, and compared AMIGA with GNAN’s source-node attribution (kjMs);
  • evaluated higher-rank interactions and feature–topology correlation (RduP);
  • added two tree-based datasets with ground-truth explanations (rHYP);
  • expanded the discussion of practical interpretability, computational cost, motif size, and scalability (rHYP).

The reviewers’ final responses confirmed the value of these revisions:

  • kjMs stated that the revisions substantially addressed the original concerns, had no further questions, and increased the rating from 2 to 3.
  • RduP stated that the new analyses adequately addressed the main concerns and maintained the original positive assessment.
  • rHYP confirmed that the concerns about interpretability and explanation quality had been addressed, increased the rating to 2, and acknowledged the method’s applicability to molecular graphs.

We thank the reviewers and the Area Chair again for their time and remain available for any further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.