Higher Order Topological Deep Learning

This direction is merged into another direction: 1-WL expressivity limit, topological expressivity, subgraph-based expressivity. Its last version stays readable and citable here.

Version 1 of 1548 papers762 conceptsfrozen 2026-10-04

Standard Graph Neural Networks (GNNs) operate on a simple premise: nodes exchange messages with their immediate neighbors. While powerful, this pairwise interaction model has a fundamental "glass ceiling." As Christopher Morris and colleagues proved in Weisfeiler and Leman Go Neural: Higher-Order Graph Neural Networks, most GNNs are no more expressive than the 1-dimensional Weisfeiler-Leman (1-WL) isomorphism test. This means they are "blind" to certain structural patterns—they cannot distinguish between a single six-carbon ring (benzene) and two disconnected three-carbon triangles.

To break through this ceiling, a new direction has emerged: Topological Deep Learning (TDL). Instead of treating data as a collection of points and lines, TDL "lifts" graphs into higher-dimensional topological spaces where triangles, cycles, and voids become first-class citizens.

The Lifting Maneuver

The core intuition of TDL is that the most important information in a complex system often resides in its multi-way interactions, not just its pairwise ones. In a survey of the field, Mathilde Papillon and Mustafa Hajij describe this as the "lifting" maneuver: preprocessing a graph to identify higher-order structures like cliques or rings and representing them as higher-dimensional "cells" in a complex.

By transforming a graph into a Simplicial Complex (made of points, lines, and triangles) or a Cell Complex (which allows for more general polygons), researchers like Claudio Battiloro enable neural networks to explicitly learn from these structures. In Cell Attention Networks, for instance, the model doesn't just aggregate neighbor nodes; it uses attention to understand how edges share a common polygon, allowing it to detect chemical motifs that standard GNNs would only infer indirectly.

Message Passing Across Dimensions

Once a graph is lifted, message passing becomes a multi-level conversation. In a standard GNN, messages only flow between nodes. In a Topological Neural Network (TNN), messages flow:

  1. Within a dimension: Edges talk to adjacent edges.
  2. Up the hierarchy: Nodes send information to the edges they form; edges send information to the faces they bound.
  3. Down the hierarchy: Faces send signals back down to their constituent edges and nodes.

Cong Liu and David Ruhe take this further in Clifford Group Equivariant Simplicial Message Passing Networks by representing these higher-order structures using Clifford algebras. By treating a triangle as a "bivector," the model gains a geometric awareness that is naturally equivariant to rotations and reflections, making it far more effective for physical and molecular tasks.

Capturing Global Shape with Persistence

While higher-order message passing captures local "blobs" of connectivity, it can still struggle with global topology—the overall "skeleton" of the data. This is where Persistent Homology (PH) comes in.

PH tracks how topological features (like holes or voids) appear and disappear as we "zoom out" from the data. Max Horn and colleagues introduced Topological Graph Neural Networks, which use a differentiable layer to feed these global "persistence" features directly into a GNN. Similarly, Bastian Rieck proved in On the Expressivity of Persistent Homology in Graph Learning that PH is at least as expressive as the $k$-WL test, allowing it to distinguish complex "strongly-regular" graphs that even advanced message-passing schemes fail to differentiate.

Why It Works: Overcoming Structural Blindness

The success of these methods stems from their ability to compute features that are mathematically impossible for a standard GNN to reach. Several researchers have proposed "tricks" to bridge this gap:

  • Subgraph Counting: Giorgos Bouritsas and Fabrizio Frasca showed that simply counting how many times a node appears in a triangle or clique and adding that count as a feature allows a 1-WL GNN to exceed its theoretical limits.
  • Homomorphism Counts: Pascal Welke demonstrated that counting mappings from "pattern graphs" into a target graph provides an "expectation-complete" representation, theoretically capable of distinguishing any two non-isomorphic graphs.
  • Labeling Tricks: Muhan Zhang proved that by labeling nodes relative to a target set, a simple GNN can learn to compute complex topological heuristics like "common neighbor counts" that are otherwise invisible to it.

By moving from pairwise links to higher-order complexes, TDL provides a mathematically rigorous way to encode the "shape" of data. As Theodore Papamarkou argues in Topological Deep Learning is the New Frontier for Relational Learning, this shift from graphs to complexes is not just a technical upgrade—it is a fundamental expansion of the types of relationships neural networks can understand.

Papers in this version

Showing 60 of 548 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_higher_order_topological_deep_learning_f32a4ad9,
  title = {Higher Order Topological Deep Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/higher-order-topological-deep-learning/rdr_f32a4ad9202840ebad2b4d6b20e80984},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_higher_order_topological_deep_learning_f32a4ad9_v1,
  title = {Higher Order Topological Deep Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/higher-order-topological-deep-learning/rdr_f32a4ad9202840ebad2b4d6b20e80984/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}