We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Node Feature Kernels Increase Graph Convolutional Network Robustness

Abstract: The robustness of the much-used Graph Convolutional Networks (GCNs) to perturbations of their input is becoming a topic of increasing importance. In this paper, the random GCN is introduced for which a random matrix theory analysis is possible. This analysis suggests that if the graph is sufficiently perturbed, or in the extreme case random, then the GCN fails to benefit from the node features. It is furthermore observed that enhancing the message passing step in GCNs by adding the node feature kernel to the adjacency matrix of the graph structure solves this problem. An empirical study of a GCN utilised for node classification on six real datasets further confirms the theoretical findings and demonstrates that perturbations of the graph structure can result in GCNs performing significantly worse than Multi-Layer Perceptrons run on the node features alone. In practice, adding a node feature kernel to the message passing of perturbed graphs results in a significant improvement of the GCN's performance, thereby rendering it more robust to graph perturbations. Our code is publicly available at:this https URL
Comments: 16 pages, 5 figures
Subjects: Machine Learning (cs.LG); Social and Information Networks (cs.SI); Machine Learning (stat.ML)
Cite as: arXiv:2109.01785 [cs.LG]
  (or arXiv:2109.01785v3 [cs.LG] for this version)

Submission history

From: Johannes Lutzeyer [view email]
[v1] Sat, 4 Sep 2021 04:20:45 GMT (869kb,D)
[v2] Thu, 20 Jan 2022 17:35:03 GMT (867kb,D)
[v3] Mon, 21 Feb 2022 15:26:20 GMT (287kb,D)

Link back to: arXiv, form interface, contact.