Structure-Guided Input Graph for GNNs facing Heterophily

Victor M. Tenorio, Madeline Navarro, Samuel Rey, Santiago Segarra, and Antonio G. Marques · 2024

Read Structure-Guided Input Graph for GNNs facing Heterophily by Victor M. Tenorio, Madeline Navarro, Samuel Rey, Santiago Segarra, and Antonio G. Marques (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-01757.txt.

Description

Victor Tenorio and co-authors rewire the input graph of graph neural networks so the models work on graphs where neighbors differ, in a 2024 paper. Such networks learn from linked data by pooling news from neighbors. They shine when neighbors share a label, but stall on the many real graphs where neighbors tend to differ.

Instead of changing the network, the team change the graph it sees. They link each node to others with like structure, picked by role traits such as degree or by global traits such as centrality. On that rebuilt graph, neighbors share labels far more often, so the same network reads smoother signals. Tests show standard models gain accuracy on heterophilic data sets when fed the new graph. The trick is cheap, since it swaps the input rather than the learner.

Catalog record

CREATORVictor M. Tenorio, Madeline Navarro, Samuel Rey, Santiago Segarra, and Antonio G. Marques
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-01757

Browse the full library · About the Tower of Babel