Uporaba metod globokega učenja za generiranje sintetičnih relacijskih podatkov

Real-world databases are predominantly relational, comprising multiple interlinked tables that contain complex structural and statistical dependencies. Learning generative models on relational data has shown great promise in generating synthetic data, which can power privacy-sensitive workloads and unlock access to previously underutilized data. However, existing methods often struggle to capture this complexity, typically reducing relational data to conditionally generated individual tables, imposing limiting structural assumptions and a fixed ordering of tables where there is none. To address these limitations, we introduce RelDiff, a novel diffusion generative method that jointly synthesizes all tables in a relational database by explicitly modeling their foreign key graph structure. RelDiff combines a joint graph-conditioned diffusion process for attribute synthesis and a graph generator based on the Stochastic Block Model for structure generation. The decomposition of graph structure and relational attributes ensures both high fidelity and referential integrity, both of which are crucial aspects of synthetic relational database generation. Experiments on 11 benchmark datasets demonstrate that RelDiff consistently outperforms state-of-the-art methods in producing realistic and coherent synthetic relational databases.