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.
Introduction
Topological Deep Learning (TDL) has gained significant traction due to its ability to model "higher-order" interactions—complex relationships involving multiple entities simultaneously, such as a group of proteins interacting in a metabolic pathway. These interactions are typically modeled using simplicial complexes, which are collections of nodes, edges, triangles, and higher-dimensional simplices.
However, a major limitation of current TDL frameworks is the assumption of symmetry. Most existing models treat these relationships as undirected, effectively "symmetrizing" the data. In reality, many systems are inherently asymmetric, such as citation networks, signal flow in the brain, or metabolic pathways. Symmetrizing these structures leads to a significant loss of information. While directionality is well-defined for edges in a graph, defining it for higher-order structures like triangles or tetrahedra is non-trivial. This paper addresses this gap by formalizing higher-order directionality and introducing Directed Simplicial Neural Networks (Dir-SNNs).
Higher-Order Topological Directionality
The core contribution of this work is the formal definition of Higher-Order Topological Directionality. Unlike undirected complexes where a simplex is a set of vertices, the authors treat simplices as ordered tuples. This ordering allows for the definition of Face Maps ($d_i$), which are operators that systematically remove the $i$-th vertex from an ordered simplex to identify its "faces."
By leveraging these face maps, the authors define two new types of neighborhood relations that preserve orientation:
- Down $(k, i, j)$-adjacency: Two $k$-simplices are neighbors if they share a common $(k-1)$-dimensional face at specific indices $i$ and $j$.
- Up $(k, i, j)$-adjacency: Two $k$-simplices are neighbors if they are both faces of the same $(k+1)$-dimensional simplex.
This vertex-ordering approach allows the model to distinguish between different types of flows, such as "equidirected" versus "circular" flows in triangles, which is impossible in undirected or symmetrized models.
Directed Simplicial Neural Networks (Dir-SNNs)
The Dir-SNN architecture utilizes these directed adjacency relations to perform message passing. In this framework, a simplex updates its hidden 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).
The authors prove that symmetrizing a directed simplicial complex is a non-injective process. This means that two distinct directed structures can become indistinguishable once their orientation is removed, highlighting the necessity of a model that can process directed information directly.
Expressivity and the D-SWL Test
To evaluate the theoretical power of Dir-SNNs, the authors introduce the Directed Simplicial Weisfeiler-Leman (D-SWL) test. This test serves as a measure of the network's ability to distinguish non-isomorphic structures. They prove that Dir-SNNs are strictly more powerful than standard Directed Graph Neural Networks (Dir-GNNs) because they can detect higher-order directed features, such as directed triangles, that node-and-edge models overlook.
In a toy discrimination task involving the graphs shown above, standard Dir-GNNs achieved only 50% accuracy (equivalent to random guessing). In contrast, Dir-SNNs achieved 100% accuracy, empirically confirming their superior expressivity.
Empirical Results
Beyond theoretical expressivity, Dir-SNNs were evaluated on a Synthetic Source Localization task. In this experiment, a signal originates at a source edge and diffuses through a network generated via a Stochastic Block Model (SBM). The goal is to identify the community of origin.
The results showed that Dir-SNNs consistently outperformed both undirected SNNs and directed GNNs across various Signal-to-Noise Ratio (SNR) levels. Furthermore, when tested on undirected complexes, Dir-SNNs performed comparably to standard SNNs, demonstrating that the directed architecture is robust and does not lose performance in the absence of directionality.
Conclusion and Future Work
Dir-SNNs represent the first framework for higher-order directionality in Topological Deep Learning. By moving beyond symmetrization, this model captures the inherent asymmetries of complex systems. While the current results are promising, future work will focus on scaling the model to larger real-world datasets and developing a spectral theory for directed simplicial complexes to enable more advanced filtering operations.
Summary updated