Structure-based graph convolutional networks with frequency filter

doi:10.1016/j.patrec.2022.11.005

Structure-based graph convolutional networks with frequency filter

2022-12-01

3

PRE

AI

DOI：10.1016/j.patrec.2022.11.005

Original

Original request for help

Save

Abstract

En 中文

Message-passing neural networks (MPNNs) have attracted a lot interesting in academia and industry and have been applied to various graph analytical problems of real-world applications and achieved prominent successes. For the most of current works on MPNNs, they mainly focus on two categories: spectral-based and spatial-based methods. The former seeks to distill useful information (e.g. low-pass and high-pass signals) and the later designs structural schemes. However, it is not enough to only utilize one of them. We tackle this drawback by proposing an efficient and elegant method of taking advantage of the structural graph neural network (GNN) for advanced spectral-based information filtering. Through the learnable frequency components and global virtual neighbor nodes, the proposed scheme is elaborated to obtain informative and structured messages of neighbor nodes and latent space correlational nodes. Experimental results show that our method outperforms the state-of-the-art performance in a series of datasets of graph tasks. (c) 2022 Elsevier B.V. All rights reserved.

Keywords:

Network representation learning

Node embeddings

Graph filtering

Graph convolution neural network

Structure-based graph convolutional networks with frequency filter

Abstract

Journal

Organization