Vpliv relacijske sheme na uspešnost grafovskih nevronskih mrež

Relational deep learning applies graph neural networks to data stored in relational databases by constructing graphs directly from database schemas. This thesis investigates how relational database schemas influence graph neural network performance, focusing on the over-squashing phenomenon that limits information propagation. We demonstrate that graphs derived from relational schemas exhibit performance patterns consistent with over-squashing theory, where models perform well with shallow networks but degrade with increased depth due to topological bottlenecks. To address these limitations, we develop two graph rewiring approaches: a computationally efficient approximation of Stochastic Discrete Ricci Flow and a schema-aware heuristic method that leverages domain knowledge about relational structures. Experimental results on two datasets show that targeted rewiring strategies can provide substantial performance improvements, suggesting that graph topology should be considered a design choice rather than a fixed constraint in relational deep learning systems.