A Simple Spectral Failure Mode for Graph Convolutional Networks

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Priebe, Carey E., Shen, Cencheng, Huang, Ningyuan, Chen, Tianyi
Format: Preprint
Published: 2020
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913404948578304
author Priebe, Carey E.
Shen, Cencheng
Huang, Ningyuan
Chen, Tianyi
author_facet Priebe, Carey E.
Shen, Cencheng
Huang, Ningyuan
Chen, Tianyi
contents Neural networks have achieved remarkable successes in machine learning tasks. This has recently been extended to graph learning using neural networks. However, there is limited theoretical work in understanding how and when they perform well, especially relative to established statistical learning techniques such as spectral embedding. In this short paper, we present a simple generative model where unsupervised graph convolutional network fails, while the adjacency spectral embedding succeeds. Specifically, unsupervised graph convolutional network is unable to look beyond the first eigenvector in certain approximately regular graphs, thus missing inference signals in non-leading eigenvectors. The phenomenon is demonstrated by visual illustrations and comprehensive simulations.
format Preprint
id arxiv_https___arxiv_org_abs_2010_13152
institution arXiv
publishDate 2020
record_format arxiv
spellingShingle A Simple Spectral Failure Mode for Graph Convolutional Networks
Priebe, Carey E.
Shen, Cencheng
Huang, Ningyuan
Chen, Tianyi
Machine Learning
Neural networks have achieved remarkable successes in machine learning tasks. This has recently been extended to graph learning using neural networks. However, there is limited theoretical work in understanding how and when they perform well, especially relative to established statistical learning techniques such as spectral embedding. In this short paper, we present a simple generative model where unsupervised graph convolutional network fails, while the adjacency spectral embedding succeeds. Specifically, unsupervised graph convolutional network is unable to look beyond the first eigenvector in certain approximately regular graphs, thus missing inference signals in non-leading eigenvectors. The phenomenon is demonstrated by visual illustrations and comprehensive simulations.
title A Simple Spectral Failure Mode for Graph Convolutional Networks
topic Machine Learning
url https://arxiv.org/abs/2010.13152