SoftEdge: Regularizing Graph Classification with Random Soft Edges

TL;DR

SoftEdge introduces a novel graph augmentation technique that assigns random weights to edges, preserving graph semantics and improving GNN performance and robustness against overfitting and depth-related issues.

Contribution

The paper proposes SoftEdge, a simple yet effective method for graph augmentation that maintains node and connectivity integrity while enhancing GNN accuracy and resilience.

Findings

SoftEdge outperforms traditional augmentation methods in accuracy.

It demonstrates robustness to GNN depth-related degradation.

Maintains original graph semantics during augmentation.

Abstract

Augmented graphs play a vital role in regularizing Graph Neural Networks (GNNs), which leverage information exchange along edges in graphs, in the form of message passing, for learning. Due to their effectiveness, simple edge and node manipulations (e.g., addition and deletion) have been widely used in graph augmentation. Nevertheless, such common augmentation techniques can dramatically change the semantics of the original graph, causing overaggressive augmentation and thus under-fitting in the GNN learning. To address this problem arising from dropping or adding graph edges and nodes, we propose SoftEdge, which assigns random weights to a portion of the edges of a given graph for augmentation. The synthetic graph generated by SoftEdge maintains the same nodes and their connectivities as the original graph, thus mitigating the semantic changes of the original graph. We empirically show that this simple method obtains superior accuracy to popular node and edge manipulation approaches and notable resilience to the accuracy degradation with the GNN depth.