We develop a novel data-driven nonlinear mixup mechanism for graph data
augmentation and present different mixup functions for sample pairs and their
labels. Mixup is a data augmentation method to create new training data by
linearly interpolating between pairs of data samples and their labels. Mixup of
graph data is challenging since the interpolation between graphs of potentially
different sizes is an ill-posed operation. Hence, a promising approach for
graph mixup is to first project the graphs onto a common latent feature space
and then explore linear and nonlinear mixup strategies in this latent space. In
this context, we propose to (i) project graphs onto the latent space of
continuous random graph models known as graphons, (ii) leverage convex
clustering in this latent space to generate nonlinear data-driven mixup
functions, and (iii) investigate the use of different mixup functions for
labels and data samples. We evaluate our graph data augmentation performance on
benchmark datasets and demonstrate that nonlinear data-driven mixup functions
can significantly improve graph classification.

本文提出了一种新的基于数据的非线性 mixup 机制，为样本对及其标签提供不同的 mixup 功能，通过在公共潜在特征空间上投射图形数据并探索非线性数据驱动的 mixup 策略，结合凸聚类生成一个最佳的数据增强方法来提高图分类的性能。