acceptodds
Under review as a conference paper at ICLR 2027

Learning Self-Explainable Edge Representations via Compositional Geometry

Abstract

Representation learning is central to graph machine learning, with link prediction as a central task. However, most graph models represent edges through opaque similarity functions, offering limited insight into how the latent memberships of two nodes combine to produce, or inhibit, a particular interaction. Many networks naturally admit a compositional view, where nodes express soft membership over multiple latent archetypes and edges may arise through different combinations of these memberships. Motivated by this structure, we propose a compositional graph model that learns self-explainable edge representations. Nodes are represented as simplex-valued compositions over learned archetypes, with each archetype defining an edge channel that combines the shared membership of a node pair in that archetype with their Aitchison agreement over the remaining archetypes. The resulting channel values jointly determine the edge probability and, after normalization, form a compositional representation of the edge itself. This yields explanations by construction: the same quantities used for prediction describe how an interaction is allocated across latent archetypes. We establish that these channels retain relational information lost by global membership overlap and full compositional distance, respond predictably to membership perturbations, and support exact counterfactual reasoning through channel restriction. Across link-prediction benchmarks, the model achieves competitive performance against strong graph-learning baselines while enabling new analyses of how latent memberships shape individual interactions. In particular, its compositional structure supports semantic membership interventions and counterfactual studies that probe which latent channels carry individual edge predictions.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
compositional data, graph representation learning, Aitchison geometry

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TNonOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors submitted a paper titled “Learning Self-Explainable Edge Representations via Compositional Geometry”

The submission includes a main PDF and a Supplementary Material zip file with code and data for replicability and reproducibility

The paper is organized in a standard structure. The AI USE STATEMENT, the Ethical statement, and the REPRODUCIBILITY STATEMENT are all included. A long appendix is also included.

The problem addressed by the authors is that traditional graph machine learning models for link prediction often rely on opaque similarity functions that fail to explain how node properties combine to form or inhibit specific interactions. To address this limitation, the authors propose a compositional graph model in which nodes express soft memberships over learned archetypes, enabling edge probabilities to be computed via self-explainable, archetype-specific channels. As a result, the proposed approach achieves competitive performance on standard link prediction benchmarks while enabling exact counterfactual reasoning and semantic membership interventions.

The experimental setting includes 5 datasets and compares against: Euclidean embeddings, ROLE2VEC(Ahmedet al., 2018), matrixfactorisation (NETMF(Qiuet al., 2018)), mixed-member ship models (MMSBM(Airoldietal.,2007), MNMF(Wangetal.,2017)) and simplex-based latent distancem odels (SLIM-RAA(Nakisetal.,2023a), HM-LDM(Nakisetal., 2022), AICOG(Nakisetal.,2026)).

Critical strengths

The paper addresses an interesting topic for the community. The approach is well formalized, and the appendix includes formal proofs. Reproducibility and replicability are considered relevant aspects of the contribution by the authors.

Critical weaknesses

I don't have major concerns only small and minor suggestions: Some references are pointing to the ArXiv editions and can be updated, as an expample two candidate updates are following:

  1. The reference to ArXiv "Mixed membership stochastic blockmodels" can be updated with:

@article{JMLR:v9:airoldi08a, author = {Edoardo M. Airoldi and David M. Blei and Stephen E. Fienberg and Eric P. Xing}, title = {Mixed Membership Stochastic Blockmodels}, journal = {Journal of Machine Learning Research}, year = {2008}, volume = {9}, number = {65}, pages = {1981--2014}, url = {http://jmlr.org/papers/v9/airoldi08a.html} }

  1. The reference to ArXiv "Aitchison’s Compositional Data Analysis 40 Years on: A Reappraisal" can be updated with:

Michael Greenacre. Eric Grunsky. John Bacon-Shone. Ionas Erb. Thomas Quinn. "Aitchison’s Compositional Data Analysis 40 Years on: A Reappraisal." Statist. Sci. 38 (3) 386 - 410, August 2023. https://doi.org/10.1214/22-STS880

and more...

Questions for authors

I don't have specific questions

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

This paper introduces ACE, a compositional graph model for self-explainable link prediction. Each node is represented by a distribution over latent archetypes, and each archetype defines an edge channel combining shared membership with agreement between complementary compositions under Aitchison geometry. Channel strengths determine edge probabilities, while their normalized values provide intrinsic edge explanations. The authors derive theoretical guarantees for targeted membership interventions and exact channel removal, and analyze the additional information retained by complementary channels. Experiments on synthetic and real networks investigate link prediction, archetype recovery, explanation sparsity, semantic alignment, and agreement between model counterfactuals and changes observed after refitting.

Critical strengths

First, the proposed ACE provides a coherent formulation in which the same archetype-specific channels determine both predictions and explanations.

Second, the evaluation part examines predictive accuracy, synthetic archetype recovery, semantic alignment, and responses to interventions.

Critical weaknesses
  1. The theoretical contributions are low-information; at least one result reduces to immediate algebra from the model definition. 3.3 follows immediately from Eq. (9), since p^(−S) decreases monotonically in W(S)=Σω_k; picking the largest ω_k is just inverting a monotone function. I find that the Thm 3.1 is only an existence statement.

  2. Figure 1 does not convey the claimed intuition and appears unfinished.

  3. The predictive gains are marginal, so the paper stands entirely on the interpretability claims — which are only weakly validated. The ablation shows the three distance-based mechanisms differ by at most 0.009 AUC on real graphs, and the authors state that the ACE/AICoG difference is "largely determined by estimation rather than the channel rule."

Questions for authors

Please see weakness.

  1. Please clarify what each theorem adds beyond Eqs. (3)–(5), and consider restructuring the presentation so the paper leads with the specific insight each result provides.

  2. Please improve the overall representation of the method.

Reviewer_GV3mOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a novel approach to representing edges as a set of channels, each having a probability of edges, with the sum of the probabilities to 1. A node is characterized by a coordinate in Aitchison space with KK archetype, and each channel represents shared membership in one of the KK architypes. While using Aitchison space as a representation basis is not entirely a new idea, the paper departs from the previous work by introducting complement distances, which are the distances between two nodes without an architype, which reflects not only the overall similarity of the archtype memberships, as well as the ratio balance of the architypes. Consequently, the proposed representation becomes more expressive than ordinal Aitchison distance. Tested their methods through internal validations as well as external validations using Cora dataset.

Critical strengths

The key strengths of the paper are the introduction of the complement distance. It shows that this distance metric is more expressive, meaning that the ordinal Aitchison distance cannot distinguish some cases, but the proposed one can. This contributes to the literature as an additional enhancement to improve the expressiveness of the representation, enabling the learning of more nuanced network structures.

Critical weaknesses
  1. While the authors identified their contribution as the ability to "interpret" the model prediction. However, in my view, it is not interpretable. I acknowledge that "interpretability" means differently within field. But let me clarify on that more speifically as follows. First, it does not provide interpretations of the predicted results. While the model provides the decomposition of probabilities into KK channel, what each channel means was guessed posteriori by humans. I'd suggest that the authors shift this claim about interpretability to something else. Even say that the decomposition itself is sufficient for interpretability, the level of interpretability is no greater than that of the mixed-membership stochastic block model. In the experiment, the proposed method showed that the model learned the topic assortatibity (nodes with the same topics are likely to be connected) and use this as a basis for superior interpretability. However, DC-MMSBM would also produce similar topically assortative clusters. Furtheremore, the interpretability is demonstrated only for one network. If the interpretability is the core claim, i'd suggest authors to test with more networks to demonstrate its interpretability.

  2. Identifiability - the authors claimed that the model is rotation invariant up to label permutation of architype, which makes sense. But it is not sufficient to say thatit is identifiable. More specifically, the following text "The remaining ambiguity is only permutation of archetype labels, as in mixed-membership models." and "Strong performance is retained with the superior identifiability of ACE." To claim the identifiability, the authors must show that the mapping from the parameters to probabilities over graph is injective. The authors need to show the identifiability proof or weaken the language.

  3. One of the key novelties of the proposed method is the introduction of the complement distance. It is, however, not clearly demonstrated the effectiveness of this key departure. Table 8 shows the alblation study comparing the ordinal vs the complement distance. The complement distance has no better or comparable at best than the full full distance (AUC-ROC; mean of five runs). Is the AICoG in the main text trained in the same procedure as ACE? If not the performance gain demonstrated is a compound of the training framework and distance metric. the authors must decouple the two by adding an extra version of AICog with the same training procedure as the ACE to demonstrate the effectiveness of the complement distance. A thought - the results led me think that in empirical networks one channel dominates so little is added by reallocating the probabilities. Indeed, Table 12 showed that a single channel is dominated.

Questions for authors
  1. What is the "interpretability" offered by ACE beyond the DC-MMSBM?
  2. Is ACE identifiable?
  3. Is the performance gain attributed to complement distance or training procedure?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.