pith. sign in

arxiv: 2605.00716 · v2 · pith:6CCSBAYXnew · submitted 2026-05-01 · 💻 cs.LG · cs.SI

Aitchison Embeddings for Learning Compositional Graph Representations

Nikolaos Nakis , Chrysoula Kosma , Panagiotis Promponas , Michail Chatzianastasis , Giannis Nikolentzos This is my paper

Pith reviewed 2026-07-01 07:41 UTC · model grok-4.3

classification 💻 cs.LG cs.SI
keywords graph embeddingsAitchison geometrycompositional representationsnode classificationlink predictioninterpretabilityILR coordinatesrole mixtures
0
0 comments X

The pith

Graph nodes represented as mixtures over latent archetypes via Aitchison geometry produce embeddings that are competitive on standard tasks and interpretable by construction.

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

The paper introduces a framework that treats nodes as compositions on the simplex, each a mixture over a small set of archetypal factors. These compositions are transformed into unconstrained Euclidean vectors through isometric log-ratio coordinates that preserve the underlying Aitchison distances. The resulting embeddings reflect relative trade-offs among the archetypes, support optimization without constraints, and remain coherent when subsets of components are removed and renormalized. Experiments show the approach matches strong baselines on node classification and link prediction while supplying explanations directly from the geometry rather than through later analysis. Both fixed and learnable bases for the coordinates are considered.

Core claim

Nodes are represented as simplex-valued compositions and embedded via isometric log-ratio (ILR) coordinates, which preserve Aitchison distances while enabling unconstrained optimization in Euclidean space. This yields intrinsically interpretable embeddings whose geometry reflects relative trade-offs among archetypes and supports coherent behavior under component restriction; we consider both fixed and learnable ILR bases.

What carries the argument

Isometric log-ratio (ILR) coordinates that map simplex compositions to Euclidean vectors while exactly preserving Aitchison distances and subcompositional coherence.

If this is right

  • Embeddings are interpretable by construction because each coordinate directly encodes relative contributions of the archetypes.
  • Subcompositional coherence permits principled removal and renormalization of component subsets while keeping distances well-defined.
  • The same geometry supports both fixed and learnable ILR bases without changing the overall optimization setup.
  • Performance remains competitive with strong baselines on node classification and link prediction across tested graphs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The subcomposition mechanism could be used to test which groups of archetypes most affect downstream predictions by systematically ablating them.
  • If real networks frequently exhibit role-mixture structure, the same Aitchison embedding could be applied to other compositional data such as topic distributions or chemical mixtures.
  • Learnable ILR bases might reveal dataset-specific archetype contrasts that are not apparent with fixed bases.

Load-bearing premise

Many networks naturally admit a role-mixture view where nodes are best described as mixtures over latent archetypal factors.

What would settle it

On a graph where node roles cannot be expressed as mixtures over a small fixed set of archetypes, the method would show clear accuracy drops relative to standard non-compositional embeddings on both node classification and link prediction.

Figures

Figures reproduced from arXiv: 2605.00716 by Chrysoula Kosma, Giannis Nikolentzos, Michail Chatzianastasis, Nikolaos Nakis, Panagiotis Promponas.

Figure 1
Figure 1. Figure 1: Overview of AICoG. Nodes are represented as compositions on the simplex and compared in Aitchison geometry. An ILR isometry maps compositions to an unconstrained Euclidean space, where distances preserve Aitchison distances. where I ∈ R (K−1)×(K−1) is the identity matrix. Given a node role zi ∈ ∆K−1 , its ILR coordinates are defined as xi = ILR(zi) = log(zi) ⊤V ∈ R K−1 , (5) where the logarithm is applied … view at source ↗
Figure 2
Figure 2. Figure 2: Cora dataset (D=8). Label-wise distributions along ILR balances under three valid bases: Helmert (left), learned (center), and varimax-rotated learned (right). Top row shows balance loadings (archetypal contributions to each log-ratio contrast); bottom row shows label-wise distributions of the corresponding ILR coordinates. −7.5 −5.0 −2.5 0.0 2.5 5.0 7.5 PC1 (20.1% var) -8.0 -6.0 -4.0 -2.0 0.0 2.0 4.0 6.0 … view at source ↗
Figure 3
Figure 3. Figure 3: Interpretable trade-off trajectories under subcomposition (Cora). Node embeddings are shown in a 2D PCA projection of ILR coordinates (Helmert basis), with PCA fit separately for each K and nodes colored by label. The overlaid curve traces a paired log-ratio intervention applied to the same node across panels, increasing archetype a and decreasing b followed by closure. The intervention is defined intrinsi… view at source ↗
Figure 3
Figure 3. Figure 3: Interpretable trade-off trajectories under subcomposition (Cora). Node embeddings are shown in a 2D PCA projection of ILR coordinates (Helmert basis), with PCA fit separately for each K and nodes colored by label. The overlaid curve traces a paired log-ratio intervention applied to the same node across panels, increasing archetype a and decreasing b followed by closure. The intervention is defined intrinsi… view at source ↗
Figure 4
Figure 4. Figure 4: Subcompositional evaluation on CORA (trained at D=64). We evaluate semantically meaningful component removal by restricting each simplex-based representation to K′ compo￾nents, applying closure, and probing the resulting D=K′ − 1 embeddings without retraining. Curves are averaged over 50 ran￾dom removal masks; higher is better in both panels. pronounced under aggressive compression. 8 [PITH_FULL_IMAGE:fig… view at source ↗
Figure 4
Figure 4. Figure 4: Subcompositional evaluation on CORA (trained at D=64). We evaluate semantically meaningful component removal by restricting each simplex-based representation to K′ compo￾nents, applying closure, and probing the resulting D=K′ − 1 embeddings without retraining. Curves are averaged over 50 ran￾dom removal masks; higher is better in both panels. ant to the choice of ILR basis: different bases yield different … view at source ↗
read the original abstract

Representation learning is central to graph machine learning, powering tasks such as link prediction and node classification. However, most graph embeddings are hard to interpret, offering limited insight into how learned features relate to graph structure. Many networks naturally admit a role-mixture view, where nodes are best described as mixtures over latent archetypal factors. Motivated by this structure, we propose a compositional graph embedding framework grounded in Aitchison geometry, the canonical geometry for comparing mixtures. Nodes are represented as simplex-valued compositions and embedded via isometric log-ratio (ILR) coordinates, which preserve Aitchison distances while enabling unconstrained optimization in Euclidean space. This yields intrinsically interpretable embeddings whose geometry reflects relative trade-offs among archetypes and supports coherent behavior under component restriction; we consider both fixed and learnable ILR bases. Across node classification and link prediction, our method achieves competitive performance with strong baselines while providing explainability by construction rather than post-hoc. Finally, subcompositional coherence enables principled component restriction: removing and renormalizing subsets preserves a well-defined geometry, which we exploit via subcompositional dimensionality removal to probe how archetype groups influence representations and predictions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper proposes a compositional graph embedding framework based on Aitchison geometry. Nodes are represented as simplex-valued compositions over latent archetypal factors and embedded using isometric log-ratio (ILR) coordinates (both fixed and learnable bases) to enable Euclidean optimization while preserving Aitchison distances and subcompositional coherence. The central empirical claim is that the resulting embeddings achieve competitive performance on node classification and link prediction relative to strong baselines, with the added property of intrinsic explainability rather than post-hoc analysis.

Significance. If the empirical competitiveness and geometric properties hold under scrutiny, the work supplies a principled inductive bias for role-mixture structures in networks. The explicit use of Aitchison geometry and ILR embeddings for compositional data, together with subcompositional dimensionality removal, constitutes a concrete strength that could improve interpretability in graph representation learning.

major comments (1)
  1. [Abstract] Abstract: the assertion of 'competitive performance with strong baselines' is load-bearing for the central claim yet is presented without any visible quantitative results, error bars, dataset names, or statistical details; this prevents verification that the performance advantage (or parity) is robust rather than an artifact of particular splits or baselines.
minor comments (2)
  1. [Abstract] The motivation paragraph treats the role-mixture view as a modeling choice rather than deriving it from graph properties; a brief justification or reference to when this inductive bias is expected to be appropriate would improve clarity.
  2. [Abstract] Notation for the ILR transformation and the distinction between fixed versus learnable bases is introduced without an explicit equation or small illustrative example, making the transition from simplex to Euclidean space harder to follow on first reading.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their thoughtful review and for highlighting the need for greater transparency in the abstract. We agree that the central empirical claim requires concrete support and will revise the abstract accordingly while preserving its brevity. No standing objections remain after addressing this point.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion of 'competitive performance with strong baselines' is load-bearing for the central claim yet is presented without any visible quantitative results, error bars, dataset names, or statistical details; this prevents verification that the performance advantage (or parity) is robust rather than an artifact of particular splits or baselines.

    Authors: We concur that the abstract should provide sufficient quantitative grounding for the competitiveness claim. In the revised version we will insert concise performance highlights (e.g., mean accuracy or AUC with standard deviation across splits), name the primary datasets (Cora, CiteSeer, PubMed, and the link-prediction benchmarks), and reference the number of runs or statistical tests used, while remaining within abstract length limits. The full tables and error-bar figures already present in the experimental section will be cross-referenced. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The derivation applies the established Aitchison geometry and isometric log-ratio (ILR) transformation—standard tools from compositional data analysis—to node representations on the simplex as an explicit modeling choice. Performance and explainability claims rest on downstream empirical evaluation against external baselines on node classification and link prediction tasks, not on any internal redefinition or self-referential fitting. The role-mixture motivation is presented as an inductive bias rather than a derived theorem, and subcompositional coherence follows directly from the properties of the ILR map without reducing predictions to the inputs by construction. No self-citation chains, fitted parameters renamed as predictions, or ansatzes smuggled via prior work appear in the provided derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review prevents exhaustive enumeration; the role-mixture assumption is treated as domain_assumption and the choice of ILR basis (fixed or learnable) introduces potential free parameters whose values are not specified here.

axioms (1)
  • domain assumption Many networks naturally admit a role-mixture view where nodes are mixtures over latent archetypal factors.
    Stated in the first paragraph of the abstract as motivation for the framework.

pith-pipeline@v0.9.1-grok · 5747 in / 1157 out tokens · 19494 ms · 2026-07-01T07:41:14.836936+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Rank Is Not Capacity: Spectral Occupancy for Latent Graph Models

    cs.LG 2026-05 unverdicted novelty 7.0

    Spectra defines and controls effective capacity in graph embeddings via the Shannon effective rank of a trace-normalized kernel spectrum, making capacity a post-fit property rather than a pre-training hyperparameter.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages · cited by 1 Pith paper · 4 internal anchors

  1. [1]

    Learning Role-based Graph Embeddings

    Ahmed, N. K., Rossi, R., Lee, J. B., Willke, T. L., Zhou, R., Kong, X., and Eldardiry, H. Learning role-based graph embeddings.arXiv preprint arXiv:1802.02896,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Mixed membership stochastic blockmodels

    URL https://arxiv.org/abs/0705.4485. Aitchison, J. The Statistical Analysis of Compositional Data.Journal of the Royal Statistical Society: Series B (Methodological), 44(2):139–160,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Explainability Techniques for Graph Convolutional Networks

    Baldassarre, F. and Azizpour, H. Explainability tech- niques for graph convolutional networks.arXiv preprint arXiv:1905.13686,

    work page internal anchor Pith review Pith/arXiv arXiv 1905

  4. [4]

    Epasto, A

    doi: 10.1023/A:1023818214614. Epasto, A. and Perozzi, B. Is a Single Embedding Enough? Learning Node Representations that Capture Multiple Social Contexts. InThe World Wide Web Conference, pp. 394–404,

    work page doi:10.1023/a:1023818214614

  5. [5]

    org/abs/2201.05197

    URL https://arxiv. org/abs/2201.05197. Grover, A. and Leskovec, J. node2vec: Scalable Feature Learning for Networks. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Dis- covery and Data Mining, pp. 855–864,

    work page arXiv

  6. [6]

    URL https: //doi.org/10.1198/016214502388618906

    doi: 10.1198/016214502388618906. URL https: //doi.org/10.1198/016214502388618906. Holland, P. W., Laskey, K. B., and Leinhardt, S. Stochastic blockmodels: First steps.Social Net- works, 5(2):109–137,

    work page doi:10.1198/016214502388618906

  7. [7]

    doi: https://doi.org/10.1016/0378-8733(83)90021-7

    ISSN 0378-8733. doi: https://doi.org/10.1016/0378-8733(83)90021-7. URL https://www.sciencedirect.com/ science/article/pii/0378873383900217. Jin, J., Ke, Z. T., and Luo, S. Mixed membership estima- tion for social networks,

    work page doi:10.1016/0378-8733(83)90021-7

  8. [8]

    org/abs/1708.07852

    URL https://arxiv. org/abs/1708.07852. Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    work page arXiv

  9. [9]

    Semi-Supervised Classification with Graph Convolutional Networks

    Kipf, T. Semi-supervised classification with graph con- volutional networks.arXiv preprint arXiv:1609.02907,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    J., Bulusu, K

    Lin, C., Sun, G. J., Bulusu, K. C., Dry, J. R., and Hernandez, M. Graph neural networks including sparse interpretabil- ity.arXiv preprint arXiv:2007.00119,

    work page arXiv 2007

  11. [11]

    Nakis, N., Celikkanat, A., Boucherie, L., Djurhuus, C., Burmester, F., Holmelund, D

    URL https://arxiv.org/abs/2206.03463. Nakis, N., Celikkanat, A., Boucherie, L., Djurhuus, C., Burmester, F., Holmelund, D. M., Frolcov ´a, M., and Mørup, M. Characterizing polarization in social networks using the signed relational latent distance model. In Ruiz, F., Dy, J., and van de Meent, J.-W. (eds.),Proceedings of The 26th International Conference o...

    work page arXiv

  12. [12]

    Perozzi, B., Al-Rfou, R., and Skiena, S

    URL https://arxiv.org/ abs/2304.10851. Perozzi, B., Al-Rfou, R., and Skiena, S. DeepWalk: On- line Learning of Social Representations. InProceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 701–710,

    work page arXiv

  13. [13]

    Don’t walk, skip! online learning of multi-scale network em- beddings

    Perozzi, B., Kulkarni, V ., Chen, H., and Skiena, S. Don’t walk, skip! online learning of multi-scale network em- beddings. InProceedings of the 2017 IEEE/ACM In- ternational Conference on Advances in Social Networks Analysis and Mining, pp. 258–265,

    2017

  14. [14]

    org/abs/2005.07959

    URL https://arxiv. org/abs/2005.07959. Rozemberczki, B., Kiss, O., and Sarkar, R. Karate Club: An API Oriented Open-source Python Framework for Unsupervised Learning on Graphs. InProceedings of the 29th ACM International Conference on Information and Knowledge Management (CIKM ’20), pp. 3125–3132. ACM,

    work page arXiv 2005

  15. [15]

    Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y ., and Liu, T.-Y

    doi: 10.1086/226141. Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y ., and Liu, T.-Y . Do transformers really perform badly for graph representation?Advances in neural information processing systems, 34:28877–28888,

    work page doi:10.1086/226141

  16. [16]

    =σ(α−g(∥x i −x j∥2)), x i ∈R k−1. Proof. Since ILR is bijective, for any collection {xi}n i=1 ⊂R k−1 there exists a unique collection {zi}n i=1 ⊂∆ k−1 ◦ such thatx i = ILR(zi)for alli, and converselyz i = ILR−1(xi)exists for alli. BecauseILRis an isometry, for alli < j, ∥xi −x j∥2 =∥ILR(z i)−ILR(z j)∥2. Substituting this identity into the respective expre...

    2008

  17. [17]

    Link prediction.For link prediction, we follow the widely adopted evaluation protocol of Perozzi et al

    For SLIM-RAA, HM-LDM, MMSBM, and SIMPLEX-EUCLIDEAN, we optimized the negative Bernoulli log-likelihood (matching AICOG up to the model-specific log-odds parameterization) using learning rate 0.05 for 5,000 epochs, so differences are attributable only to the log-odds form. Link prediction.For link prediction, we follow the widely adopted evaluation protoco...

    2014