Overcoming Information Bottlenecks in Directed Graph Neural Networks through Rewiring

Yu He, Ishani Karmarkar, Ellen Vitercik

Keywords: graph neural networks, directed graphs, graph rewiring

TL;DR: We propose a rewiring framework for directed graphs that alleviates over-squashing and strengthens long-range dependencies in directed GNNs.

Abstract: Graph Neural Networks (GNNs) have become an essential tool in relational learning. However, most research has focused on undirected graphs, overlooking the directional structure that shapes many real-world systems. Directionality encodes important asymmetries in domains such as traffic routing, social networks, and causal modeling, where ignoring edge orientation may lose semantic information and hinder predictive performance. At the same time, message passing in GNNs is limited by over-squashing and long-range dependency, challenges that can be alleviated by modifying the graph topology. While recent rewiring techniques based on spectral graph theory have shown promise for undirected graphs, they do not extend naturally to directed settings. In this work, we introduce a rewiring framework for directed graphs that improves information flow while preserving inherent asymmetries and reachability constraints. Our method directly tackles over-squashing and long-range dependency issues, advancing graph rewiring into the directed regime and broadening the use of GNNs in critical real-world tasks.

Submission Type: Extended abstract (max 4 main pages).

Submission Number: 112