Tensor Graph Convolutional Network for Dynamic Graph Representation Learning

Ling Wang,Ye Yuan

2024-01-13

Abstract:Dynamic graphs (DG) describe dynamic interactions between entities in many practical scenarios. Most existing DG representation learning models combine graph convolutional network and sequence neural network, which model spatial-temporal dependencies through two different types of neural networks. However, this hybrid design cannot well capture the spatial-temporal continuity of a DG. In this paper, we propose a tensor graph convolutional network to learn DG representations in one convolution framework based on the tensor product with the following two-fold ideas: a) representing the information of DG by tensor form; b) adopting tensor product to design a tensor graph convolutional network modeling spatial-temporal feature simultaneously. Experiments on real-world DG datasets demonstrate that our model obtains state-of-the-art performance.

Machine Learning,Artificial Intelligence

What problem does this paper attempt to address?

This paper attempts to solve the spatio - temporal continuity problem in dynamic graph (DG) representation learning. Specifically, most of the existing DG representation learning models combine the graph convolutional network (GCN) and the sequence neural network to model spatio - temporal dependencies through two different types of neural networks. However, this hybrid design fails to capture the spatio - temporal continuity of DG well.

To solve the above problems, the author proposes a tensor graph convolutional network (TGCN) based on tensor product to model spatio - temporal features simultaneously in a convolutional framework. The main contributions of this method include:

1. **Information Representation**: Represent the information of DG in tensor form and embed node information to obtain the latent feature tensor.
2. **Tensor Graph Convolution Model**: Propose a tensor graph convolution model based on tensor product to model spatio - temporal dependencies simultaneously in a convolutional framework, in which a learnable mixing matrix is designed to capture different temporal - dependency scales between graph snapshots.
3. **Experimental Verification**: Conduct extensive experiments on real - world DG datasets to verify the effectiveness of the proposed TGCN model. The experimental results show that TGCN outperforms the existing state - of - the - art models in accurately estimating link weight values.

Through these improvements, TGCN can capture spatio - temporal dependencies in dynamic graphs more effectively, thereby improving the performance of DG representation learning.