An Empirical Evaluation of Rewiring Approaches in Graph Neural Networks

TL;DR

This paper systematically evaluates graph rewiring methods in GNNs using a training-free setting, revealing that such rewiring rarely improves performance on real-world tasks, thus questioning their practical utility.

Contribution

It introduces a training-free evaluation framework for graph rewiring methods, enabling clearer assessment of their actual benefits in GNNs.

Findings

Rewiring rarely improves node classification performance.

Training-free evaluation isolates effects of rewiring from training issues.

Most rewiring methods show limited practical benefits on real-world data.

Abstract

Graph neural networks compute node representations by performing multiple message-passing steps that consist in local aggregations of node features. Having deep models that can leverage longer-range interactions between nodes is hindered by the issues of over-smoothing and over-squashing. In particular, the latter is attributed to the graph topology which guides the message-passing, causing a node representation to become insensitive to information contained at distant nodes. Many graph rewiring methods have been proposed to remedy or mitigate this problem. However, properly evaluating the benefits of these methods is made difficult by the coupling of over-squashing with other issues strictly related to model training, such as vanishing gradients. Therefore, we propose an evaluation setting based on message-passing models that do not require training to compute node and graph representations. We perform a systematic experimental comparison on real-world node and graph classification tasks, showing that rewiring the underlying graph rarely does confer a practical benefit for message-passing.