Abstract:Geometric deep learning (GDL) models have demonstrated a great potential for the analysis of non-Euclidian data. They are developed to incorporate the geometric and topological information of non-Euclidian data into the end-to-end deep learning architectures. Motivated by the recent success of discrete Ricci curvature in graph neural network (GNNs), we propose TorGNN, an analytic Torsion enhanced Graph Neural Network model. The essential idea is to characterize graph local structures with an analytic torsion based weight formula. Mathematically, analytic torsion is a topological invariant that can distinguish spaces which are homotopy equivalent but not homeomorphic. In our TorGNN, for each edge, a corresponding local simplicial complex is identified, then the analytic torsion (for this local simplicial complex) is calculated, and further used as a weight (for this edge) in message-passing process. Our TorGNN model is validated on link prediction tasks from sixteen different types of networks and node classification tasks from three types of networks. It has been found that our TorGNN can achieve superior performance on both tasks, and outperform various state-of-the-art models. This demonstrates that analytic torsion is a highly efficient topological invariant in the characterization of graph structures and can significantly boost the performance of GNNs.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: How to better incorporate geometric and topological information in Graph Neural Networks (GNNs) to improve the performance of GNNs when dealing with non - Euclidean data. Specifically, the authors propose a new graph neural network model - TorGNN (Torsion enhanced Graph Neural Network), which enhances the ability of GNNs to capture local features of graph structures by introducing analytic torsion. ### Background and Motivation In recent years, Geometric Deep Learning (GDL) models have shown great potential in non - Euclidean data analysis. These models aim to incorporate the geometric and topological information of non - Euclidean data into end - to - end deep learning architectures. Although GNNs have achieved remarkable success in many tasks, how to effectively combine geometric and topological information remains a key issue. ### The Role of Analytic Torsion Analytic torsion is a topological invariant that can distinguish homotopically equivalent but not topologically equivalent spaces. The authors use analytic torsion to characterize the local structure of a graph and use it as an edge weight in the message - passing process. Specifically, for each edge, the corresponding local simplicial complex is identified, its analytic torsion is calculated, and it is used as the weight of this edge in the message - passing. ### Model Advantages 1. **Better Local Structure Capture**: By introducing analytic torsion, TorGNN can more accurately capture the local complexity of graph data. 2. **Surpass Existing Models**: Experimental results show that TorGNN outperforms several existing advanced models in both link prediction and node classification tasks. 3. **Wide Applicability**: TorGNN performs well on different types of networks (such as biomedical networks, social networks, cooperation networks, Internet peer - to - peer networks, etc.), showing strong generalization ability. ### Experimental Verification The authors conducted experiments on multiple datasets, including 16 different types of networks for link prediction tasks and 3 networks for node classification tasks. The experimental results show that TorGNN is significantly superior to other models in terms of AUC and AUPR metrics, especially when dealing with large - scale and dense networks. ### Formula Presentation The definition of analytic torsion is as follows: \[ \log T(K)=\frac{1}{2}\sum_{p = 0}^{n}(- 1)^{p}p\zeta'_p(0)\] where \(\zeta'_p(0)\) is the derivative of the Zeta function at \(s = 0\), and \(\zeta_p(s)\) is defined as: \[ \zeta_p(s)=\sum_{\lambda_i>0}\frac{1}{\lambda_i^s}\] ### Conclusion By introducing analytic torsion, TorGNN not only improves the performance of GNNs, but also provides a new and effective tool for geometric deep learning, especially showing significant advantages when dealing with complex graph structures.

Torsion Graph Neural Networks

Tensor-view Topological Graph Neural Network

MGNN: Graph Neural Networks Inspired by Distance Geometry Problem

Efficient Topology-aware Data Augmentation for High-Degree Graph Neural Networks

Tensorized Hypergraph Neural Networks

TopoTune : A Framework for Generalized Combinatorial Complex Neural Networks

Graph Coordinates and Conventional Neural Networks -- An Alternative for Graph Neural Networks

Adaptive curvature exploration geometric graph neural network

TANGNN: a Concise, Scalable and Effective Graph Neural Networks with Top-m Attention Mechanism for Graph Representation Learning

TF-GNN: Graph Neural Networks in TensorFlow

Isometric Graph Neural Networks

Modeling Graphs Beyond Hyperbolic: Graph Neural Networks in Symmetric Positive Definite Matrices

Conditional Uncertainty Quantification for Tensorized Topological Neural Networks

Enhancing the Resilience of Graph Neural Networks to Topological Perturbations in Sparse Graphs

Dirac--Bianconi Graph Neural Networks -- Enabling Non-Diffusive Long-Range Graph Predictions

Graph Neural Tangent Kernel: Fusing Graph Neural Networks with Graph Kernels

Graph neural networks and non-commuting operators

Graph Neural Networks Inspired by Classical Iterative Algorithms

T-HyperGNNs: Hypergraph Neural Networks via Tensor Representations

Orthogonal Graph Neural Networks

E(n) Equivariant Topological Neural Networks