Is Graph Mixup Beneficial? Investigating Interpolation And Empirical Performance of Graph Mixup Methods

Simon Forbat, Rainer Gemulla

Keywords: Graph Neural Networks, Data Augmentation, Mixup, Graph Classification, Graph Edit Distance, Representation Learning, Evaluation

TL;DR: We investigate graph mixup and analyze prediction performance as well as interpolation properties of prior methods empirically.

Abstract:

Mixup is a widely used data augmentation technique that constructs new training
examples by interpolating between existing ones. While simple and effective in
domains like vision and language, applying mixup to graph data is non-trivial
and independent empirical evidence for its effectiveness is lacking. To fill this
gap, we conducted an independent evaluation following established evaluation
protocols for graph classification and found that none of the state-of-the-art mixup
methods yielded statistically significant improvements over the no-mixup baseline.
To obtain further insights, we analyzed the graphs generated from existing mixup
methods from an interpolation perspective using the graph edit distance. We found
that (i) many mixup methods failed to interpolate well, (ii) high interpolation error
led to performance degradation, and (iii) even optimal interpolation did not lead to
performance improvements. Our findings highlight the need for a more rigorous
exploration and evaluation of mixup for graphs.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 16401