Učenje grafovskih predstavitev za evalvacijo sintetičnih relacijskih podatkov

Evaluating the utility of synthetic relational databases is challenging, as existing approaches rely on manual feature engineering or single-table flattening, which obscure relational structure and reduce scalability. This thesis introduces RDL-utility, a general framework that represents relational databases as heterogeneous graphs and trains graph neural networks (GNNs) directly on the graphs. Using a standardized AutoComplete task, RDL-utility measures how well models trained on synthetic data perform on real held-out data. Experiments on five real-world databases, including an ablation study across six GNN architectures, show that diffusion-based generative methods achieve the highest utility, although no single method consistently outperforms all others. RDL-utility provides a reproducible, structure-sensitive evaluation, establishing a foundation for future research and applications.