Application of Partition of Unity Methods in the Graph Neural Networks’ framework

Graph Neural Networks (GNNs) have emerged as powerful tools for node classification tasks in graph-structured data, where deep learning approaches currently represent the state of the art. However, a major limitation of these models lies in the substantial computational resources required to process large-scale graphs. To address this issue, this work investigates several graph partitioning techniques aimed at decomposing a large graph into smaller, more manageable subgraphs. Each subgraph is used to train an independent model, and the individual outputs are then combined through an ensemble strategy. This partition-based approach, when combined with Backpropagation-Free Graph Neural Networks, results in a significant reduction in computational cost while maintaining competitive performance levels.