A systematic approach to random data augmentation on graph neural networks

TL;DR

This paper introduces a comprehensive framework for random data augmentation in graph neural networks, enabling systematic comparison, theoretical universality, and the development of improved augmentation methods.

Contribution

It presents a unified framework capturing all previous RDA techniques, proves their universality, and offers a method for automatic training and comparison of RDAs.

Findings

The framework is universally applicable under natural conditions.

Automatically trained RDAs outperform existing methods.

New RDAs derived from the framework improve state-of-the-art results.

Abstract

Random data augmentations (RDAs) are state of the art regarding practical graph neural networks that are provably universal. There is great diversity regarding terminology, methodology, benchmarks, and evaluation metrics used among existing RDAs. Not only does this make it increasingly difficult for practitioners to decide which technique to apply to a given problem, but it also stands in the way of systematic improvements. We propose a new comprehensive framework that captures all previous RDA techniques. On the theoretical side, among other results, we formally prove that under natural conditions all instantiations of our framework are universal. On the practical side, we develop a method to systematically and automatically train RDAs. This in turn enables us to impartially and objectively compare all existing RDAs. New RDAs naturally emerge from our approach, and our experiments demonstrate that they improve the state of the art.