Higher-Order Interaction Modeling with Simplicial Complexes

149 concepts122 papers1 proposal

Graphs are the bedrock of relational machine learning. Nodes represent entities, and edges represent the relationships between them. But this binary framing forces a massive compromise: it reduces all group interactions to a collection of independent pairs.

Consider a three-person research collaboration. In a standard graph, this is represented as a triangle of three edges. But does that triangle represent one cohesive paper written by all three authors together, or three separate papers written in pairs? A standard graph cannot tell the difference. As noted by Alice Patania et al. in The shape of collaborations, reducing multi-agent interactions to simple pairs loses the true "shape" and cohesion of group dynamics.

To capture these multi-way dependencies, a growing body of research shifts from traditional graphs to simplicial complexes. As argued by Gunnar E. Carlsson et al. in Data science for graphs, this topological framework allows us to model many-body relationships natively, preserving group synergies that are otherwise lost.

The Anatomy of a Simplicial Complex

What exactly is a simplicial complex? Think of it as a hierarchical toolkit for data. Instead of just points and lines, we build with higher-dimensional geometric primitives called simplices:

  • A 0-simplex is a vertex (a single node).
  • A 1-simplex is an edge (a pair of nodes).
  • A 2-simplex is a filled triangle (a triplet of nodes).
  • A 3-simplex is a solid tetrahedron (a quadruplet of nodes).

A simplicial complex is simply a collection of these simplices glued together under a strict rule known as downward closure. As Christian Bick explains in What Are Higher-Order Networks?, downward closure means that if a set of nodes forms a higher-order interaction (like a 2-simplex triangle), all of its subsets (the constituent edges and vertices) must also exist in the complex.

  Graph Representation           Simplicial Complex
     (Pairwise Only)          (Higher-Order Interaction)

          Node A                       Node A
          /    \                       / Filled \
         /      \                     /  2-Simp  \
     Edge 1    Edge 2             Edge 1  (ABC)  Edge 2
       /          \                 /              \
      /            \               /                \
  Node B -------- Node C       Node B -------- Node C
          Edge 3                       Edge 3

This structure allows us to explicitly differentiate between a hollow triangle (three independent pairwise relationships) and a filled triangle (a genuine three-way interaction). This distinction is critical across domains. For instance, in wireless networks, Michael Robinson et al. show in Analyzing wireless communication network vulnerability with homological invariants how higher-dimensional faces represent mutual interference among groups of nodes, capturing resource contention that pairwise models miss. In structural biology, JunJie Wee et al. demonstrate in A cohomology-based Gromov-Hausdorff metric approach for quantifying molecular similarity that modeling molecular systems as simplicial complexes mathematically captures the internal geometry of rings and cavities, which traditional atom-bond graphs overlook.

Topological Deep Learning via Message Passing

How do neural networks operate on these multi-dimensional spaces? Standard Graph Neural Networks (GNNs) update node representations by aggregating features from immediate neighbors across edges. However, GNNs hit a fundamental theoretical wall: they cannot distinguish certain non-isomorphic graph structures, a limitation bounded by the Weisfeiler-Lehman (1-WL) isomorphism test.

To break through this barrier, Cristian Bodnar et al. introduced Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks (MPSNs). Instead of restricting message passing to a single dimension, MPSNs allow information to flow hierarchically across different ranks of simplices. An edge can send a message "up" to a triangle, and a triangle can send a message "down" to its boundary edges. By lifting graphs into simplicial complexes, the authors prove that this multi-dimensional message passing is strictly more expressive than standard GNNs, successfully distinguishing complex structures that even the powerful 3-WL test fails to resolve.

Similarly, frameworks like Simplicial Neural Networks, developed by Michaël Defferrard et al. in Simplicial Neural Networks, leverage these topological structures to perform convolutions directly on higher-order relations. This multi-way awareness dramatically improves tasks like missing data imputation and short-text classification, where group-wise dependencies provide vital context, as seen in Contrastive Learning with Simplicial Convolutional Networks for Short-Text Classification.

Tracking Flows with the Hodge Laplacian

In many real-world networks, the data of interest does not live on the nodes—it lives on the edges or higher-dimensional faces. Think of traffic flowing through a city grid, ocean currents, or trajectories moving through space. Standard graph signal processing struggles with these because it ignores the geometric orientation and rotational properties of the space.

By representing environments as simplicial complexes, we gain access to Hodge theory, a powerful branch of differential geometry. Just as the standard graph Laplacian helps us understand smooth signals across nodes, the Simplicial Laplacian allows us to process signals on edges and higher-order faces.

As demonstrated by Elvin Isufi et al. in Finite Impulse Response Filters for Simplicial Complexes, the Simplicial Laplacian enables the Hodge decomposition, which mathematically splits any edge flow into three orthogonal components:

  1. Gradient flow: Flows that originate at a source and end at a sink (irrotational).
  2. Curl flow: Flows that circulate around a higher-dimensional cavity or triangle (rotational).
  3. Harmonic flow: Flows that pass through the network without a clear source or sink, often navigating around structural holes.
     Gradient Flow                Curl Flow               Harmonic Flow
    (Source to Sink)            (Rotational)            (Around a Hole)

     O --> O --> O               O ------> O               O --> O --> O
                 |               ^         |               |   | H |   |
                 v               |         v               |   | O |   v
                 O               O <------ O               O   | L |   O
                                                           |   | E |   ^
                                                           v   |   |   |
                                                           O --> O --> O

This decomposition provides an elegant mechanism for filtering and prediction. For example, Florian Frantzen et al. in Outlier Detection for Trajectories via Flow-embeddings project movement paths into harmonic spaces to detect outlier trajectories that take unusual topological detours around obstacles. Meanwhile, Jae Choi et al. show in SNN-PDE: Learning Dynamic PDEs from Data with Simplicial Neural Networks how incorporating these higher-order geometric structures allows models to learn partial differential equations (PDEs) on irregular spatial grids, capturing surface-level dynamics that standard pairwise models completely miss.

Generalizing Beyond Simplices

While simplicial complexes offer a rigorous foundation, the field is expanding into even more flexible topological representations. One limitation of simplicial complexes is their rigid combinatorial structure—every $k$-dimensional shape must be broken down into triangles (triangulation).

To achieve greater representational efficiency, researchers are moving toward cell complexes. As discussed by Danilo Comminiello et al. in Semantic Communications Based on Adaptive Generative Models and Information Bottleneck and Soledad Villar et al. in The Second Learning on Graphs Conference: Preface, cell complexes allow for arbitrary polygons (like squares or pentagons) instead of strictly triangles. This flexibility enables models to capture complex semantic relations and edge flows more sparsely, reducing parameter overhead while maintaining high reconstruction accuracy.

Furthermore, to handle the directed nature of real-world interactions, Manuel Lecha et al. introduce semi-simplicial sets in Directed Semi-Simplicial Learning with Applications to Brain Activity Decoding, allowing multiple distinct directed motifs to exist over the same set of nodes. Ultimately, as unified frameworks like Copresheaf Topological Neural Networks (Mustafa Hajij et al. in Copresheaf Topological Neural Networks: A Generalized Deep Learning Framework) emerge, topology is proving to be a necessary evolution for deep learning—transforming how we represent the deep, multi-dimensional geometry of interconnected data.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 121 papers
  2. Version 1 · 2026-10-04 · 200 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_simplicial_complex_representation_higher_248bc572,
  title = {simplicial complex representation (higher-order interaction modeling)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/simplicial-complex-representation-higher-order-interaction-modeling/rdr_248bc5722a6e4901a7640ad7d5ff64da},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 2, frozen text):

@misc{lacuna_simplicial_complex_representation_higher_248bc572_v2,
  title = {simplicial complex representation (higher-order interaction modeling)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/simplicial-complex-representation-higher-order-interaction-modeling/rdr_248bc5722a6e4901a7640ad7d5ff64da/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake