Graph Representation Learning for Evaluation of Synthetic Relational Data

Jurkovič, Martin (author), Šubelj, Lovro (mentor)

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.

• Language: English
• Keywords: synthetic data, relational databases, data generation, graph representation learning, graph neural networks, empirical comparison, data quality evaluation, utility
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2025
• PID: 20.500.12556/RUL-173771
• COBISS.SI-ID: 254172931
• Publication date in RUL: 22.09.2025

Secondary language

• Language: Slovenian
• Title: Učenje grafovskih predstavitev za evalvacijo sintetičnih relacijskih podatkov

Abstract

Ocenjevanje uporabnosti sintetičnih relacijskih baz podatkov je zahtevno, saj obstoječi pristopi temeljijo na ročnem ustvarjanju značilk ali združevanju tabel v eno samo, kar zakriva relacijsko strukturo in zmanjšuje razširljivost. Ta magistrska naloga uvaja RDL-utility, splošni pristop, ki relacijske baze podatkov pretvori v heterogene grafe in na njih neposredno trenira grafovske nevronske mreže (GNN). Z uporabo standardizirane prediktivne naloge AutoComplete, RDL-utility meri, kako dobro modeli, naučeni na sintetičnih podatkih, delujejo na resničnih, ločenih testnih podatkih. Eksperimenti na petih realnih podatkovnih bazah, vključno s študijo preko šestih GNN arhitektur, kažejo, da generativni pristopi, ki bazirajo na difuziji, dosegajo najvišjo uporabnost, čeprav noben posamezen pristop dosledno ne prekaša vseh drugih. RDL-utility zagotavlja reproducibilno, na strukturo občutljivo ocenjevanje ter postavlja temelje za prihodnje raziskave in aplikacije.

• Keywords: sintetični podatki, relacijske baze, generiranje podatkov, učenje grafovskih predstavitev, grafovske nevronske mreže, empirična primerjava, evalvacija kvalitete podatkov, uporabnost