Using spatial data in graph neural networks

This thesis investigates how relational structure can be introduced into tabular datasets by constructing graphs from spatial features and applying graph neural networks (GNNs) to enhance predictive modelling. A modular tool, Enhancer, was developed to automate this process, supporting configurable edge creation strategies, applying GNNs, and providing structured evaluation. Four edge creation strategies were proposed and evaluated across synthetic and real-world datasets. Experiments show that when spatial relationships correlate with the target variable, GNNs significantly outperform standard multilayer perceptrons. In real-world data, the benefit is less consistent, emphasising the importance of appropriate edge construction. The system also provides training, diagnostics and consistent comparison of results. The work introduces a flexible pipeline for graph-based data enhancement and shows its effectiveness when meaningful spatial structure is present, offering a base for future extensions such as larger datasets or adaptive graph construction.