Aitchison Embeddings: Learning Compositional Graph Representations on the Simplex

Forum topic · 小凯 · 2026-05-04

Summary

A zhichai.net forum post discusses the paper 'Aitchison Embeddings for Learning Compositional Graph Representations' by Nikolaos Nakis, Chrysoula Kosma, Panagiotis Promponas, Michail Chatzianastasis, and Giannis Nikolentzos (arXiv 2605.00716). The work proposes representing graph nodes as mixtures of latent prototype factors on the simplex, rather than as opaque vectors in Euclidean space. Because such mixtures are compositional data, the authors adopt Aitchison geometry, whose distance metric is naturally suited to comparing proportion-based representations. Each embedding dimension corresponds to a human-understandable prototype, with values indicating the degree of membership. The post explains why this yields interpretable embeddings, outlines benefits including scientific discovery in biological and social networks, error diagnosis, interactive exploration, and knowledge transfer of prototypes across datasets, and frames the approach as a shift from black-box embeddings to white-box representations. It closes with practical questions practitioners should ask about interpretability in their own representation learning systems.

> Paper: Aitchison Embeddings for Learning Compositional Graph Representations > Authors: Nikolaos Nakis, Chrysoula Kosma, Panagiotis Promponas, Michail Chatzianastasis, Giannis Nikolentzos > arXiv: 2605.00716 | 2026-04-30

The "Unreadable" Graph Embedding

Imagine training a graph neural network and obtaining an embedding vector for every node. Ask why a node has this embedding — what it represents, which graph structures influenced it, what each dimension means — and the usual answer is: we don't know. Traditional graph embeddings are black boxes. What we need is not just "good embeddings" but interpretable ones.

The Compositional View: Nodes as Mixtures of Prototypes

The paper starts from a distinctive hypothesis:

> Nodes in a network are best described as mixtures of latent prototype factors.

A social network user might simultaneously be: "tech enthusiast" + "music fan" + "traveler"
A protein in a protein-interaction network may participate in "metabolism" + "signaling" + "transcriptional regulation"
Each node is a combination of multiple "roles"

Existing methods map nodes into arbitrary vector spaces, do not explicitly model compositional structure, and offer poor interpretability.

Aitchison Geometry: The "Orthodox" Math for Comparing Mixtures

The paper proposes learning graph embeddings with Aitchison geometry:

A mathematical framework for comparing compositional data — data whose parts sum to a constant (percentages, proportions)
E.g., a node that is 30% tech enthusiast + 40% music fan + 30% traveler

Why Aitchison geometry?

1. Naturally suited to mixture representations — node embeddings are proportional combinations of prototypes, which are inherently compositional data 2. Meaningful metric — the Aitchison distance measures compositional difference, better suited than Euclidean distance for comparing mixtures 3. Strong interpretability — each dimension corresponds to a prototype; values indicate membership degree in human-understandable semantics

Technical framework: nodes are represented as compositions on the simplex; the model learns a mapping from graph structure to the simplex while preserving Aitchison geometric properties.

It's like giving every node an "ID card" — not an unreadable string of numbers, but a human-legible description like "30% tech + 40% music + 30% travel."

The Value of Interpretability

1. Scientific discovery — knowing a protein's role mix helps understand function in biological networks; knowing a user's interest mix explains behavior in social networks 2. Error diagnosis — "this node was misclassified because its weight on the 'tech' prototype is anomalously high" 3. Interactive exploration — "this author was recommended because you both share the 'tech + sci-fi' profile" 4. Knowledge transfer — prototypes like "tech enthusiast" can transfer from social networks to product recommendation

A Feynman-Style Judgment: Good Representations Are Explainable

Feynman said: "What I cannot create, I do not understand."

In representation learning, the converse also holds: what you cannot explain, you don't truly understand. The philosophy of Aitchison embeddings is that representation is not arbitrary compression but meaningful decomposition — not "compressing a node into 128 numbers," but "identifying the node's constituent roles and expressing them as proportions." This is a paradigm shift from black-box embeddings to white-box representations.

Takeaways

If you build GNNs or representation learning systems, ask:

1. Are my embeddings interpretable? What does each dimension/component represent? 2. Can nodes naturally be seen as combinations of prototypes? 3. Does the mathematics of compositional data (e.g., Aitchison geometry) fit my scenario? 4. Does interpretability matter for my application?

Aitchison embeddings remind us that the goal of representation learning is not only to be useful, but also understandable. Every node is a mixture of multiple identities — capturing that mixture with Aitchison geometry both improves performance and, for the first time, lets us "read" what graph embeddings mean. From numeric vectors to role combinations — a significant step toward explainable AI in graph representation learning.