SUBRead: Clustering Sub-Graphs for Graph-Level Readout

Sitthichart Tohmuang, Mohammad Fard, Pier Marzocca, Haytham M. Fayek

Keywords: Graph Neural Networks, Graph Readout, Subgraph Clustering, Graph Representation Learning, Mode Shape Classification, Structural Dynamics

TL;DR: We present SUBRead, a readout method that combines clustering and attention to form subgraphs for graph-level representation learning on benchmarks and a real-world automotive engineering problem.

Abstract: Graph Neural Networks (GNNs) have transformed graph representation learning tasks across domains from bioinformatics and social networks to engineering applications. In graph classification, the readout function is an important component of the GNN architecture as it aggregates node features into a compact graph-level representation. Standard readouts such as sum, mean, and max often fail on complex graphs, as they cannot capture structural dependencies or contextual relationships among nodes due to the over-compressive nature of these functions, which leads to information loss. To address these challenges, we propose SUBRead, an expressive readout function which integrates subgraph clustering with attention-based weighting to produce a graph-level representation that preserves local structural information while capturing global dependencies. SUBRead is fully differentiable and compatible with various GNN architectures. Experiments on bioinformatics and social network benchmarks demonstrate that SUBRead consistently outperforms existing readouts, improving accuracy and interpretability. We further evaluate SUBRead on a real-world automotive engineering problem, where the task is to classify vibration responses of structures, referred to as structural mode shapes, using attributed graphs derived from simulation results. Unlike common graph benchmarks where graphs vary in topology, the mode shape graphs share a similar topology but significantly differ in node features, making sub-graph essential and providing a unique benchmark for readouts. The analysis demonstrates that SUBRead not only outperforms existing readouts but also provides meaningful substructures comparable to expert reasoning.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 10544