Paper Review: G-Mixup: Graph Data Augmentation for Graph Classification

In this quick review, we’ll talk about G-Mixup, a novel data augmentation mechanism that improves the generalizability of Graph Neural Networks (GNNs) by applying Mixup to estimated graphons.

Graphs G = (V, E), where V is the set of nodes/vertices in the graph, and E is the set of edges, or connections, between these nodes. Photo by JJ Ying on Unsplash.

Overview

Tags: Graph Neural Networks, Data Augmentation, Supervised Machine Learning, Graphs

Year: 2022 (AirXv), Conference/Journal TBD

Research Gap(s) Filled:

1. Extending the benefits of improved generalization through Mixup [2] to the graph domain.
2. Develops interpolation for graphs by interpolation of graph generators, i.e. graphons [1]. Allows for sampling from interpolated graphons.

Links:

1. AirXv
2. OpenReview

Abridged Summary:

Historically, data augmentation mechanisms for graphs have been limited to within-graph data augmentations, e.g. modifying nodes and edges between nodes. This paper develops the notion of between-graph data augmentation through interpolation between graphons, or classes of graphs [1].

Mixup [2], a data augmentation technique designed to improve the robustness of supervised learning approaches, has historically been applied to…