acceptodds
Under review as a conference paper at ICLR 2027

HARMONIA: Interpretable Graph Learning through Mixtures of Neural Bases

14 views

Abstract

Existing interpretable graph additive models still face limitations in either computational scalability or modeling flexibility. In terms of structural modeling, previous approaches either face quadratic scaling costs or sacrifice explicit source-to-target contribution decomposition. In terms of feature components, they rely either on per-feature neural networks or on single shared bases with limited feature specialization. We address both problems by introducing HARMONIA: Interpretable Graph Learning through Mixtures of Neural Bases, an interpretable-by-design framework. For feature modeling, HARMONIA introduces a Mixture of Neural Bases (MoNB), which routes features to specialized basis experts, enabling parameter sharing without sacrificing feature-specific specialization. For structural modeling, HARMONIA uses Relative Random Walk Probabilities (RRWP) to capture multi-hop and multi-path relationships, and proposes Sparse RRWP Aggregation (SRA) to compute these interactions through sparse graph propagation without quadratic pairwise complexity. HARMONIA retains a simple additive form in which predictions decompose into feature responses modulated by structural influence. Empirically, HARMONIA achieves stronger explanation recovery than existing interpretable graph baselines while maintaining competitive predictive performance and scaling to graphs with millions of nodes. These results show that interpretable graph learning can remain both faithful and scalable without sacrificing predictive effectiveness.

Primary area:
interpretability and explainable AI
Keywords:
Graph Learning, Graph Neural Networks, Generalized Additive Models, Graph Interpretability

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_5mfeOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces a new interpretable graph additive model that combines mixture of neural bases for feature-wise nonlinear modeling with sparse RRWP aggregation for structural propagation. MoNB routes each feature to a small subset of shared basis experts using feature identity rather than the current input, preserving globally inspectable univariate functions. SRA applies feature specific, nonnegative mixtures of random walk powers without constructing dense node pair tensors. This method is evaluated on various node and graph level tasks

Critical strengths
  1. The model provides exact feature, source node, and walk length contributions to its logits without requiring an auxiliary explainer.
  2. Shared neural bases limit per-feature parameter overhead, while exact sparse propagation avoids quadratic node-pair materialization.
Critical weaknesses
  1. Insufficiently established methodological novelty

a. Restricting RRWP to a linear topology function yields a feature-wise polynomial graph filter evaluated through sparse recursion. The authors should distinguish the novelty of its additive parameterization from this computational reformulation. A comparison against an additive neural-basis model with learned sparse polynomial propagation is needed to substantiate the contribution.

b. The integration is largely modular: MoNB supplies feature responses, while SRA propagates them without explicitly exploiting the individual expert decomposition. Moreover, the MoNB theory sets aside graph propagation, leaving unclear when its specialization advantage translates into better graph predictions. Parameter-matched comparisons against a global NBM with identical sparse aggregation are needed to establish the benefit of this particular pairing beyond the independent advantages of its components.

  1. The combination of random-walk propagation and sum pooling introduces an exact expressivity limitation. On undirected regular graphs, the random-walk matrix is doubly stochastic, so 1⊤Mt=1⊤\mathbf{1}^\top M^t = \mathbf{1}^\top. Together with ∑tθt,k=1\sum_t \theta_{t,k} = 1, the graph readout in Appendix G.6 reduces to ℓG,c=βc+∑kwk,c∑ifk(xi,k)\ell_{G,c} = \beta_c + \sum_k w_{k,c} \sum_i f_k(x_{i,k}), eliminating dependence on connectivity for fixed input features. Consequently, the model cannot distinguish regular graphs with identical node-feature multisets, even with nonconstant attributes, and neither additional experts nor propagation depths resolves this limitation. Although this does not invalidate the node-classification model, the authors should characterize this boundary and evaluate its practical significance through topology-sensitive graph-classification tests.

  2. The combined equations imply that the model uses same feature response function fk(x)f_k(x) across all propagation hops. Only the hop weights change as gk,t(x)=θt,kfk(x)g_{k,t}(x) = \theta_{t,k} f_k(x). Consequently, the model cannot learn a linear effect from a node's own feature and a quadratic effect from its neighbours for the same feature. Adding more neural-basis experts does not resolve this limitation, since their outputs are combined before graph propagation. The authors should discuss this restriction and compare against hop-specific response functions or expert-specific propagation.

Questions for authors

Please refer Weaknesses

7 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for the careful and technically detailed assessment. We would like to clarify that HARMONIA is designed to improve functional flexibility, interpretability, and scalability within a deliberately structured graph additive model, rather than to maximize unrestricted graph expressivity. Several observations raised by the reviewer follow directly from these design choices. We address each concern below.

W1a. On the novelty of SRA and polynomial graph propagation

We acknowledge that restricting RRWP to a linear topology function yields a feature-wise polynomial graph filter, and we do not claim novelty in polynomial filtering or sparse matrix multiplication themselves.

Specifically, the linearity of RRWP allows dense source-to-target aggregation to be reorganized into recursive sparse propagation without approximation or changes to the underlying predictive function. This reduces aggregation complexity from O(Tn2d)O(Tn^2d) to O(T∣E∣d)O(T|E|d), without materializing dense RRWP operators.

The methodological distinction is therefore the additive parameterization and its integration with specialized feature responses, which enables fine-grained attribution within a scalable graph learning framework, rather than the introduction of a new propagation primitive.

We will clarify this distinction and further discuss its relationship to existing polynomial graph propagation methods in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal by Authors

W1b. On the modular integration of MoNB and SRA

The modular integration of MoNB and SRA is an intentional design choice to preserve transparent and interpretable feature modeling and structural propagation. Rather than tightly coupling these mechanisms, HARMONIA separates two complementary questions: how each feature contributes to a prediction and how that contribution propagates through the graph. MoNB learns specialized nonlinear feature responses, while SRA models their propagation across source nodes and multiple hops.

Importantly, this modularity does not discard individual expert contributions. Since SRA is linear and MoNB routing depends on feature identity, propagating a mixture of expert responses is mathematically equivalent to propagating each expert response separately and combining them with the same routing weights. Thus, expert-level contributions remain analytically decomposable without requiring separate propagation for every expert.

Regarding theory, we isolate MoNB specialization from graph propagation to characterize its approximation advantages independently of graph topology. Our specialization–routing analysis identifies when specialized basis spaces improve upon a fixed-dimensional global basis space. We acknowledge that better feature approximation does not automatically guarantee better graph predictions, since propagation and supervision may suppress differences between feature representations.

To directly evaluate the benefit of specialization in graph learning, we provide parameter-matched comparisons between MoNB and a global NBM using identical sparse aggregation architectures and training protocols. Test accuracy is reported as mean ±\pm SD across seeds.

Setting (seeds)Parameters MoNB / NBMMoNB Acc. (%)Global NBM Acc. (%)
64-node synthetic (10)1,669 / 1,66383.29±0.43\mathbf{83.29\pm0.43}77.41±0.4077.41\pm0.40
Cora (5)47,641 / 47,64279.60±1.30\mathbf{79.60\pm1.30}76.87±1.6876.87\pm1.68
PubMed (5)11,878 / 11,87667.50±3.24\mathbf{67.50\pm3.24}66.90±3.4066.90\pm3.40
ogbn-arxiv (5)10,416 / 10,41780.14±2.30\mathbf{80.14\pm2.30}78.40±2.6978.40\pm2.69
IGB-Small (5)47,187 / 47,18868.93±1.08\mathbf{68.93\pm1.08}67.43±1.2667.43\pm1.26

MoNB achieves higher mean accuracy across all five settings under matched parameter budgets and identical propagation mechanisms. These results provide empirical evidence that feature specialization can benefit downstream graph prediction, consistent with the specialization advantage characterized in Theorem 1.

AuthorsOpenReview ↗

Rebuttal by Authors

W2. On the expressivity limitation for regular graphs

We agree with the reviewer's mathematical observation. For undirected regular graphs, the random-walk matrix is doubly stochastic. Consequently, under sum pooling, the normalized propagation operator preserves the global sum of each feature response, making the specified graph-level readout independent of connectivity for fixed node-feature multisets.

This is a direct consequence of combining stochastic propagation with linear sum pooling, rather than an overlooked limitation of MoNB. Increasing the number of experts or propagation depths does not resolve this property, as the reviewer correctly notes.

However, it is important to distinguish this graph-level restriction from node-level structural modeling. Although the global sum is preserved, propagation can still redistribute feature information across individual nodes according to graph connectivity. Consequently, node-level predictions remain sensitive to structural relationships.

HARMONIA is not intended to distinguish every pair of non-isomorphic graphs. Rather, it prioritizes fine-grained interpretability and efficient multi-hop structural modeling, particularly for large-scale node-level learning.

We will explicitly characterize this graph-level expressivity boundary. To assess the practical value of propagation under the same sum readout, we compared zero-hop MoNB with MoNB using 8 walk orders on three graph classification benchmarks.

DatasetHARMONIA Acc. (%)Zero-hop MoNB Acc. (%)Illustrative gain (pp)
Mutagenicity74.7070.4+4.3
PROTEINS70.7668.5+2.3
DD75.3872.3+3.1

These results show that multi-hop propagation can improve graph-level prediction while preserving HARMONIA's additive readout.

Reviewer_GK4xOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents HARMONIA, an interpretable graph additive model combining mixtures of neural bases (MoNB) for feature specialization with sparse random-walk aggregation (SRA) for scalable structural modeling. The framework enables exact feature-, source-node-, and hop-level logit decompositions. Experiments evaluate predictive performance, explanation recovery, and scalability.

Critical strengths
  1. The additive formulation provides fine-grained, model-faithful explanations through explicit feature-, source-node-, and walk-length contributions. The synthetic recovery experiments further support its interpretability.

  2. The sparse aggregation scheme avoids materializing dense node-pair matrices while preserving multi-hop structural information, enabling application to large-scale graphs with competitive predictive performance.

Critical weaknesses
  1. The novelty and individual contributions of the proposed components could be better established. In particular, MoNB lacks comparisons against capacity-matched global basis models, while SRA appears closely related to standard sparse polynomial graph propagation.

  2. The empirical evidence for interpretability and scalability remains incomplete. Explanation recovery is primarily evaluated on synthetic data whose generating mechanism aligns with the proposed model, while real-world explanation analyses are limited. Moreover, scalability is mainly supported by complexity analysis and OOM comparisons, without controlled runtime or peak-memory measurements.

Questions for authors
  1. How does MoNB compare with a wider global NBM under matched parameter budgets, and what distinguishes SRA from existing sparse polynomial propagation approaches?

  2. Could the authors provide controlled runtime and peak-memory comparisons to quantify the practical efficiency gains?

  3. How can the faithfulness and stability of HARMONIA's explanations be quantitatively evaluated on real-world graphs without ground-truth explanations?

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Q1, also W1. How does MoNB compare with a wider global NBM at a matched parameter budget, and what distinguishes SRA from standard sparse polynomial propagation?

We thank the reviewer for this question. HARMONIA addresses two key challenges in interpretable graph learning: efficiently modeling heterogeneous feature responses and capturing multi-hop structural influences while preserving fine-grained interpretability and scalability. MoNB and SRA address these complementary challenges within a unified additive framework.

Regarding MoNB, its contribution is not merely increasing the number of neural bases, but enabling feature-dependent specialization through sparse expert routing. Unlike a global NBM, which represents all feature responses within a single shared basis space, MoNB allows heterogeneous features to access specialized functional spaces while retaining parameter sharing. Our theoretical analysis characterizes this benefit through the specialization gap and establishes when specialization gains outweigh routing errors. We agree that capacity-matched global NBM comparisons would further strengthen the empirical evidence. Within each setting, MoNB and the global NBM use the same propagation architecture and training protocol. Test accuracy is mean ±\pm SD across seeds.

Setting (seeds)Parameters MoNB / NBMMoNB Acc. (%)Global NBM Acc. (%)
64-node synthetic (10)1,669 / 1,66383.29±0.43\mathbf{83.29\pm0.43}77.41±0.4077.41\pm0.40
Cora (5)47,641 / 47,64279.60±1.30\mathbf{79.60\pm1.30}76.87±1.6876.87\pm1.68
PubMed (5)11,878 / 11,87667.50±3.24\mathbf{67.50\pm3.24}66.90±3.4066.90\pm3.40
ogbn-arxiv (5)10,416 / 10,41780.14±2.30\mathbf{80.14\pm2.30}78.40±2.6978.40\pm2.69
IGB-Small (5)47,187 / 47,18868.93±1.08\mathbf{68.93\pm1.08}67.43±1.2667.43\pm1.26

These results show that MoNB consistently outperforms the global NBM under matched parameter budgets, supporting Theorem 1, which explains how feature specialization can improve approximation when its benefits outweigh routing errors.

Regarding SRA, we agree that its computation is mathematically a special case of polynomial graph propagation. We do not claim novelty in polynomial filtering or sparse matrix multiplication themselves. Rather, our key observation is that this structure is particularly well suited to resolving a fundamental challenge in interpretable graph additive models: preserving exact, fine-grained structural attribution without incurring quadratic pairwise computation.

Specifically, the linearity of RRWP aggregation allows us to reorganize dense source-to-target interactions into recursive sparse propagation without approximation or loss of feature–source-node–hop decomposition. This reduces aggregation complexity from (O(Tn2d)O(Tn^2d)) to (O(T∣E∣d)O(T|E|d)), while preserving the original interpretable formulation.

To our knowledge, prior interpretable graph additive models have not explicitly exploited this equivalence to achieve both exact RRWP-based source-to-target, hop-resolved attribution and sparse scalable computation within the same framework. Thus, the contribution of SRA lies not in inventing a new propagation primitive, but in recognizing and leveraging an existing mathematical structure to overcome a previously unresolved combination of interpretability and scalability requirements.

Thank you for suggesting the matched-budget comparison and quantitative explanation diagnostics.

AuthorsOpenReview ↗

Rebuttal by Authors

Q2, also W2. How do training and inference time and peak GPU memory compare under a controlled setup?

We profiled 100 training epochs on one RTX 4090 in FP32 with seed 0. Training time is the total synchronized compute time, excluding preprocessing, data transfers, validation, and I/O. Inference time is averaged over five synchronized forward passes over the complete test split. Memory is peak GPU allocation by the process in each phase.

Model or backendTraining Time (s)Training Peak (MiB)Inference Time (ms)Inference Peak (MiB)
Cora (2,708 nodes, full graph)
HARMONIA (SRA)0.481863.7128.7
HARMONIA (dense RRWP)3.921,47720.11537
G-NAMRFF9.921,89817.5591.5
GNAN76.552,359137.04222
ogbn-arxiv (169,343 nodes, full graph)
HARMONIA (SRA)17.001,03423.90259.3
G-NAMRFF22.201,69038.64422.3
GNANOOM during dense input allocation
IGB-Small (1,000,000 nodes, batched)
HARMONIA (SRA)231.90169275.2236.7
G-NAMRFF165.95224131.6283.4
GNANOOM during dense input allocation

This numerically verifies the implementation against Theorem 10, which proves that SRA computes the same linear RRWP operator as the dense backend. The sparse backend also delivered a 1.31×1.31\times training-compute speedup and reduced peak training memory by 38.8%. These measurements show that SRA improves efficiency while preserving the predictor and its additive decomposition.

Q3, also W2. How can we quantify explanation faithfulness and stability on real-world graphs without ground-truth attributions?

Proposition 4 gives exact logit reconstruction. To test the rankings at a fixed checkpoint, replace equal-size top, bottom, or random sets of features or source nodes with a common reference and measure the change in held-out logit margins. Edge interventions require recomputing the normalized walk operator.

On Cora, we ranked features by validation margins, removed 20% of the features, and retrained over five matched seeds (Section H.7). The random row averages five masks.

Cora retraining conditionTest accuracy (%)Change (pp)
All features79.480.000.00
Remove top-ranked features69.12−10.36-10.36
Remove bottom-ranked features80.06+0.58+0.58
Remove random features78.42−1.06-1.06

The larger accuracy drop after removing top-ranked features supports the predictive relevance of HARMONIA’s feature rankings on Cora, even after retraining on the remaining features. For stability, we compare rankings on the same validation nodes across checkpoints. The Cora five-seed summary gave mean pairwise Spearman ρ=0.77\rho=0.77 and top-20% Jaccard overlap of 0.47 over the ten seed pairs, complementing Fig. 12.

Thank you for the detailed rebuttal. The parameter-matched MoNB comparisons, runtime and memory profiling, and explanation stability analyses have substantially addressed my concerns.

I appreciate the clarification of SRA's contribution and the additional empirical evidence. For the revision, I only suggest clarifying the novelty positioning, appropriately qualifying the efficiency and faithfulness claims, and explaining the different Cora baseline accuracies.

No further experiments are necessary from my perspective. I maintain my positive assessment of the submission.

AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for the constructive follow-up and positive assessment. We are glad that the additional comparisons and analyses went a long way toward addressing their concerns.

The revised manuscript will clarify what MoNB and SRA each contribute. For SRA, we will explain how the sparse computation preserves exact structural attributions. We will also limit our efficiency claims to the settings we evaluated. The revision will describe exact additive decomposition as a property of the model and explain what our experiments show about explanation faithfulness and stability. We will also specify which evaluation protocol was used for each reported Cora baseline accuracy.

Reviewer_q68nOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HARMONIA, an interpretable graph additive framework combining a Mixture of Neural Bases with Sparse RRWP Aggregation. The feature module routes features to specialized shared basis experts, improving flexibility while retaining parameter sharing. The structural module uses normalized random-walk probabilities to model multiple propagation depths and evaluates their effects through sparse graph operations. This preserves an exact logit decomposition across features, source nodes, and walk lengths. Theory addresses specialization, routing errors, structural identifiability, and computational complexity. Experiments on node and graph classification benchmarks, synthetic mechanism recovery tasks, and ablations assess prediction, explanation quality, and scalability on large graphs.

Critical strengths

First, the paper combines feature specialization through shared neural-basis experts with efficient structural aggregation. The formulation preserves an exact decomposition of logits across features, source nodes, and propagation depths.

Then, the evaluation includes node and graph classification, large graph benchmarks, and controlled synthetic recovery tasks.

Critical weaknesses

The additive structure limits feature interactions. The routing in MORB depends only on feature identity. For a fixed graph, the logits remain additive across input features. The model therefore cannot directly represent interactions in which the effect of one feature changes with another feature’s value. The paper could make this limitation more explicit and clarify which dependencies the current formulation can capture.

Questions for authors
  1. Since MoNB routing depends only on feature identity, does it primarily improve the approximation of individual univariate functions? How would HARMONIA handle a target involving the product of two input features? Could a limited set of interaction terms be introduced while retaining a clear contribution decomposition?

  2. What additional assumptions on feature signals and supervision are needed to recover hop weights during joint training? For example, if a feature response is constant across all nodes, every propagation depth produces the same response, so different normalized hop weights yield the same aggregation. How does this ambiguity affect the interpretation of learned hop contributions?

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Q1, also the weakness. Does feature-ID routing mainly improve univariate approximation? Can HARMONIA model products of input features or add a few interaction terms while keeping their contributions interpretable?

Thank you for this question. Feature-ID routing in MoNB primarily enables feature-dependent specialization of univariate response functions while maintaining parameter sharing. In its current additive formulation, HARMONIA does not explicitly represent arbitrary cross-feature interactions, such as products of two distinct input features.

This restriction is deliberate. HARMONIA focuses on modeling nonlinear individual feature effects and their structural propagation while preserving exact feature-, source-node-, and hop-level attribution. Introducing unrestricted pairwise interactions could require up to O(d2)\mathcal{O}(d^2) components and a more complex attribution structure, particularly for high-dimensional node features.

Nevertheless, sparse interaction terms could be incorporated as separately attributable components without abandoning interpretability. For example, selected feature pairs could be modeled through additional bivariate response functions, with their interaction contributions distinguished from individual feature effects. Such an extension would increase expressivity while preserving an exact decomposition at the component level, although attribution to individual features would require additional conventions.

We will clarify this expressivity boundary in the revised manuscript and identify sparse interaction-aware modeling as a potential extension.

Q2. What feature signals and supervision are needed to identify hop weights during joint training? How should we interpret hop contributions when different weights yield the same output, as with constant feature responses?

Thank you for raising this question. HARMONIA does not assume unconditional identifiability of learned hop weights. In fact, this issue is explicitly studied in Appendix C.2. Theorem 7 establishes that normalized hop weights are uniquely recoverable from the structural operator if and only if the corresponding random-walk powers are linearly independent. Corollaries 2–3 further characterize this condition through the minimal polynomial and graph spectrum.

The reviewer's constant-signal example is a natural degenerate case. When zk=a1z_k=a\mathbf{1}, stochastic propagation gives Mtzk=zkM^t z_k=z_k for every hop. Consequently, different normalized hop weights produce the same output. This follows directly from random-walk propagation and does not contradict our operator-level identifiability result.

More generally, recovering hop weights requires (1) informative feature signals and (2) sufficient supervision. For a fixed feature signal zkz_k, normalized hop weights are uniquely identifiable from the fully observed propagated response if and only if the T−1T-1 vectors (Mt−I)zk(M^t-I)z_k, for t=1,…,T−1t=1,\ldots,T-1, are linearly independent. During joint training, supervision must additionally distinguish these hop-dependent effects, without allowing changes in learned feature responses or other parameters to compensate for changes in hop weights.

Importantly, exact decomposition and unique identifiability are different properties. HARMONIA's feature–node–hop decomposition remains algebraically exact for the learned model. However, when several hop-weight configurations produce identical observations, their individual hop contributions cannot be uniquely recovered from the data. In such cases, hop-level attribution should be interpreted cautiously, while the aggregate prediction remains well-defined.

We will clarify these signal- and supervision-dependent conditions in Appendix C to make the scope of our existing identifiability analysis explicit.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you again for your feedback. Before the discussion closes, we’d appreciate hearing whether the clarifications in our previous response addressed your concerns. If any concerns remain, we’d be happy to discuss them further.

Comments on the paper

2 comments

Dear Reviewers,

The authors have now responded to the questions and concerns raised by everyone. Could I kindly ask you to review the authors' responses, take a look at the comments from the other reviewers, and actively participate in the discussion?

Thank you all very much for your valuable time and contributions!

Best regards,

The AC

AuthorsOpenReview ↗

Official Comment by Authors

Dear Area Chair and Reviewers,

We sincerely thank you for your thoughtful reviews and constructive suggestions. In our paper, we introduce HARMONIA, an interpretable graph additive model that combines two components. The Mixture of Neural Bases (MoNB) lets different features draw on specialized shared bases while retaining one inspectable response curve per feature. Sparse RRWP Aggregation (SRA) propagates these responses across multiple walk lengths without materializing dense node-pair matrices.

We appreciate that all three reviewers highlighted HARMONIA’s exact logit decomposition by feature, source node, and walk length (q68n, GK4x, 5mfe). The discussion focused on the contributions of MoNB and SRA, the empirical evidence for efficiency and explanation quality, and the scope of the model’s expressivity and identifiability. We summarize the main clarifications and results below.

  1. Feature specialization at matched parameter budgets (GK4x, 5mfe). The Mixture of Neural Bases (MoNB) lets features draw on different subsets of shared bases while retaining one inspectable response curve per feature. We compared MoNB with a global neural basis model at matched parameter counts, using the same sparse propagation and training protocol. MoNB achieved higher mean test accuracy in all five settings: the controlled synthetic task, Cora, PubMed, ogbn-arxiv, and IGB-Small.
  2. SRA and practical efficiency (GK4x, 5mfe). We clarified how Sparse RRWP Aggregation (SRA) makes exact structural attribution practical on large sparse graphs. The sparse recursion computes the same multi-hop aggregation as the dense implementation without materializing dense node-pair matrices. On Cora, SRA reduced training compute time from 3.92 s to 0.48 s, peak training GPU memory from 1,477 MiB to 186 MiB, and inference latency from 20.11 ms to 3.71 ms. On batched IGB-Small, SRA used less peak GPU memory than G-NAMRFF and took longer in both training and inference.
  3. Expressivity of the additive model and graph readout (q68n, 5mfe). We respectfully disagree that the limitations raised in the discussion undermine HARMONIA’s methodological contribution. Related graph learning methods face similar limits when they make comparable design choices. Our goal is to learn more flexible univariate feature responses while keeping the model interpretable by design and scalable to large graphs. Sharing each feature’s response across walk lengths gives us one curve to inspect and lets us trace its contributions through the graph. Direct sum pooling can discard connectivity information on undirected regular graphs, while pooled contributions can depend on connectivity when node degrees vary. In our graph classification ablation, using eight walk orders with the same sum readout improved accuracy over zero-hop MoNB by 4.3, 2.3, and 3.1 percentage points on Mutagenicity, PROTEINS, and DD, respectively.
  4. Identifiability of hop weights (q68n). Theorem 7 characterizes when the structural operator uniquely determines the hop weights through linear independence of the random-walk powers. Recovering those weights from data also requires informative feature signals and sufficient supervision. For example, a constant feature response produces the same propagated signal at every hop. The logit decomposition remains exact for the fitted model even when multiple hop-weight configurations fit the data. We appreciate GK4x’s follow-up noting that the additional comparisons and analyses substantially addressed their concerns. In the revised manuscript, we will include these results, clarify the respective contributions of MoNB and SRA, specify the scope of our expressivity and identifiability claims, and document the evaluation protocol for each reported Cora baseline accuracy.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.