GNN-VPA: A Variance-Preserving Aggregation Strategy for Graph Neural Networks

Lisa Schneckenreiter, Richard Freinschlag, Florian Sestak, Johannes Brandstetter, Günter Klambauer +1 more
3/7/2024
cs.LGcs.AIstat.ML

No official code found. Best implementation: ml-jku/GNN-VPA (10 stars, MIT, last commit 2024-03-20).

Abstract

Graph neural networks (GNNs), and especially message-passing neural networks, excel in various domains such as physics, drug discovery, and molecular modeling. The expressivity of GNNs with respect to their ability to discriminate non-isomorphic graphs critically depends on the functions employed for message aggregation and graph-level readout. By applying signal propagation theory, we propose a variance-preserving aggregation function (VPA) that maintains expressivity, but yields improved forward and backward dynamics. Experiments demonstrate that VPA leads to increased predictive performance for popular GNN architectures as well as improved learning dynamics. Our results could pave the way towards normalizer-free or self-normalizing GNNs.

View on arXivView PDF

Code Implementations(1)

ml-jku/GNN-VPACommunity100%
MIT

A Variance-Preserving Aggregation Strategy for Graph Neural Networks

101PythonFeb 26, 20242 years agoMIT

Similar Papers

Generalised f-Mean Aggregation for Graph Neural Networks

86% match2023

Caterpillar GNN: Replacing Message Passing with Efficient Aggregation

85% match2025

Sound Logical Explanations for Mean Aggregation Graph Neural Networks

84% match2025

Graph Neural Aggregation-diffusion with Metastability

84% match2024

Position: Message-passing and spectral GNNs are two sides of the same coin

84% match2026

Entropy Aware Message Passing in Graph Neural Networks

84% match1 repo52024

Cite this paper

@article{schneckenreiter2024gnnvpa,
  title  = {GNN-VPA: A Variance-Preserving Aggregation Strategy for Graph Neural Networks},
  author = {Lisa Schneckenreiter and Richard Freinschlag and Florian Sestak and Johannes Brandstetter and Günter Klambauer and Andreas Mayr},
  year   = {2024},
  eprint = {2403.04747},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2403.04747v1}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2403.04747.svg)](https://codeofpaper.com/paper/2403.04747)

Discussion