Tensor-view Topological Graph Neural Network

Tao Wen,Elynn Chen,Yuzhou Chen

2024-01-30

Abstract:Graph classification is an important learning task for graph-structured data. Graph neural networks (GNNs) have recently gained growing attention in graph learning and have shown significant improvements in many important graph problems. Despite their state-of-the-art performances, existing GNNs only use local information from a very limited neighborhood around each node, suffering from loss of multi-modal information and overheads of excessive computation. To address these issues, we propose a novel Tensor-view Topological Graph Neural Network (TTG-NN), a class of simple yet effective topological deep learning built upon persistent homology, graph convolution, and tensor operations. This new method incorporates tensor learning to simultaneously capture Tensor-view Topological (TT), as well as Tensor-view Graph (TG) structural information on both local and global levels. Computationally, to fully exploit graph topology and structure, we propose two flexible TT and TG representation learning modules that disentangle feature tensor aggregation and transformation and learn to preserve multi-modal structure with less computation. Theoretically, we derive high probability bounds on both the out-of-sample and in-sample mean squared approximation errors for our proposed Tensor Transformation Layer (TTL). Real data experiments show that the proposed TTG-NN outperforms 20 state-of-the-art methods on various graph benchmarks.

Machine Learning,Artificial Intelligence

What problem does this paper attempt to address?

The paper aims to address the limitations of Graph Neural Networks (GNNs) in handling graph-structured data, particularly the issue that existing GNNs only utilize very limited local information around nodes, leading to multimodal information loss and excessive computational overhead. To tackle these problems, the authors propose a novel Tensor Perspective Topological Graph Neural Network (TTG-NN). This new approach combines persistent homology, graph convolution, and tensor operations, capturing both local and global structural information simultaneously by introducing tensor learning. Additionally, to fully leverage the topological structure of the graph, the method designs two flexible representation learning modules (TT and TG) to separate the aggregation and transformation of feature tensors, learning to preserve multimodal structures and reduce computational costs. Experimental results show that the proposed TTG-NN outperforms 20 state-of-the-art methods in various graph classification benchmarks, demonstrating its superior classification performance and high computational efficiency.

Tensor-view Topological Graph Neural Network

Graph Transformer Networks Based Text Representation.

Torsion Graph Neural Networks

Conditional Uncertainty Quantification for Tensorized Topological Neural Networks

Graph Transformer: Learning Better Representations for Graph Neural Networks.

Tensorized Hypergraph Neural Networks

ATPGNN: Reconstruction of Neighborhood in Graph Neural Networks with Attention-Based Topological Patterns

Topology Optimization Based Graph Convolutional Network.

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

Multi-View Tensor Graph Neural Networks Through Reinforced Aggregation

DS-TAGCN: A Dual-Stream Topology Attentive GCN for Node Classification in Dynamic Graphs.

GTAGCN: Generalized Topology Adaptive Graph Convolutional Networks

Topological Neural Networks go Persistent, Equivariant, and Continuous

Topology-Informed Graph Transformer

Hypergraph Transformer Neural Networks

Temporal Graph Transformer for Dynamic Network

TopoTune : A Framework for Generalized Combinatorial Complex Neural Networks

Graph topology enhancement for text classification

Topology-guided Hypergraph Transformer Network: Unveiling Structural Insights for Improved Representation

TGNN: A Joint Semi-supervised Framework for Graph-level Classification