Higher-Order Topological Directionality and Directed Simplicial Neural Networks

Manuel Lecha, Andrea Cavallo, Francesca Dominici, Elvin Isufi, Claudio Battiloro

TL;DR

This paper introduces the first framework for Directed Simplicial Neural Networks (Dir-SNNs), moving beyond traditional models that forcibly symmetrize relationships in higher-order data. By defining a novel notion of "higher-order directionality" based on vertex ordering and face maps, the authors create a model that is theoretically more expressive than directed graph neural networks (Dir-GNNs) and significantly better at distinguishing complex directed structures.

What Problem Does This Paper Address?

Topological Deep Learning (TDL) has gained traction because it can model "higher-order" interactions—think of a group of three proteins interacting simultaneously, rather than just three separate pairs of interactions. Usually, these are modeled as simplicial complexes (sets of nodes, edges, triangles, etc.).

However, most existing TDL models assume these relationships are undirected. In the real world, many systems are inherently asymmetric (e.g., metabolic pathways, citation networks, or signal flow in the brain). Simply ignoring the "arrows" and symmetrizing the data leads to a loss of information. The challenge is that "directionality" for a triangle or a tetrahedron is much harder to define than for a simple edge. This paper fills that gap by formalizing how direction works in higher-order spaces and building a neural network to exploit it.

Methods

The authors' core contribution is a formal definition of Higher-Order Topological Directionality. They achieve this using Face Maps ($d_i$): operators that systematically remove the $i$-th vertex from an ordered simplex to identify its "faces."

Based on these maps, they define two new types of neighborhood relations:

  1. Down $(k, i, j)$-adjacency: Two simplices are neighbors if they share a common lower-dimensional face.
  2. Up $(k, i, j)$-adjacency: Two simplices are neighbors if they are both faces of the same higher-dimensional simplex.

Because these adjacencies depend on the specific indices ($i, j$) of the vertices removed, they preserve the orientation of the structure.

The Dir-SNN architecture then uses these relations to perform message passing. A simplex (like an edge or triangle) updates its state by aggregating messages from its "up" and "down" neighbors, as well as its boundary (facets) and coboundary (the higher-order structures it belongs to).

Datasets

The paper evaluates the model on two main fronts:

  • Synthetic Source Localization: A task where a signal "spikes" at a source edge and diffuses through a network. The goal is to identify which community the signal originated from. The authors used a Stochastic Block Model (SBM) to generate 70-node graphs, which were then "lifted" into directed flag complexes.
  • Expressivity Toy Dataset: A pair of non-isomorphic directed graphs (shown in Figure 8) that are specifically designed to be indistinguishable by standard Directed GNNs but distinguishable when viewed as directed simplicial complexes.

Key Insights and Ideas

  • Symmetrization is Information Loss: The authors prove that "symmetrizing" a directed simplicial complex (treating it as undirected) is a non-injective process. Two different directed structures can look identical once you remove the arrows, making them impossible for standard models to tell apart.
  • The Power of Vertex Ordering: By treating simplices as ordered tuples of vertices rather than sets, the model can distinguish between "equidirected" flows and "circular" flows in higher-order structures (like triangles).
  • D-SWL Test: They introduce the Directed Simplicial Weisfeiler-Leman (D-SWL) test, a theoretical tool to measure the "IQ" of these networks. They prove that Dir-SNNs are strictly more powerful than Dir-GNNs because they can "see" higher-order directed features (like directed triangles) that node-and-edge models miss.

Key Results

  • Superior Expressivity: In the toy discrimination task, Dir-GNNs achieved only 50% accuracy (random guessing), while Dir-SNNs achieved 100% accuracy, confirming they can distinguish structures that graph-based models cannot.
  • Robustness to Noise: In the source localization task, Dir-SNNs consistently outperformed both undirected SNNs and directed GNNs across all Signal-to-Noise Ratio (SNR) levels.
  • Generalization: When tested on undirected complexes, Dir-SNNs performed comparably to standard SNNs, suggesting that the directed architecture doesn't "break" when the directionality is absent; it simply specializes when directionality is present.

Limitations and Future Work

  • Computational Complexity: Defining multiple types of $(i, j)$-adjacencies increases the number of messages passed per layer, which could lead to higher computational overhead compared to simpler GNNs.
  • Dataset Scarcity: The experiments are currently limited to synthetic data. The authors acknowledge the need to evaluate Dir-SNNs on a wider range of real-world datasets (e.g., biological or citation networks) where higher-order directionality is suspected to play a role.
  • Spectral Theory: The authors suggest that a future direction is developing a "spectral theory" for directed simplicial complexes, which would allow for more sophisticated filtering operations in the frequency domain, similar to Graph Fourier Transforms.

Summary updated

Abstract

Topological Deep Learning (TDL) has emerged as a paradigm to process and learn from signals defined on higher-order combinatorial topological spaces, such as simplicial or cell complexes. Although many complex systems have an asymmetric relational structure, most TDL models forcibly symmetrize these relationships. In this paper, we first introduce a novel notion of higher-order directionality and we then design Directed Simplicial Neural Networks (Dir-SNNs) based on it. Dir-SNNs are message-passing networks operating on directed simplicial complexes able to leverage directed and possibly asymmetric interactions among the simplices. To our knowledge, this is the first TDL model using a notion of higher-order directionality. We theoretically and empirically prove that Dir-SNNs are more expressive than their directed graph counterpart in distinguishing isomorphic directed graphs. Experiments on a synthetic source localization task demonstrate that Dir-SNNs outperform undirected SNNs when the underlying complex is directed, and perform comparably when the underlying complex is undirected.
BibTeX
@inproceedings{lecha2025higher,
  author = {Manuel Lecha and Andrea Cavallo and Francesca Dominici and Elvin Isufi and Claudio Battiloro},
  title = {{Higher-Order Topological Directionality and Directed Simplicial Neural Networks}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2025},
  doi = {10.1109/icassp49660.2025.10890264},
}

Loading related content...

Figure