Synthetic Relational Priors (SRP) for Generating Realistic Databases to Enhance Relational Foundation Model Pretraining

Privacy restrictions make large-scale access to real-world relational databases difficult, creating a critical data bottleneck for pretraining Relational Foundation Models (RFMs). While synthetic generators like PLUREL offer a workaround by creating databases from scratch, their simplified treatment of foreign-key connectivity, temporal dynamics, and missing values fails to capture the complex structural and distributional patterns of real datasets. To address this, we propose Synthetic Relational Priors (SRP), a novel framework for synthesizing relational databases that jointly models graph geometry and data distributions. SRP captures heterogeneous foreign-key connectivity using degree-corrected popularity parameters, models irregular temporal dynamics through continuous event-time sampling, and generates informative missingness using mixed mechanisms. Our results confirm that SRP successfully replicates real-world heavy-tailed topological structures, bursty temporal patterns, and diverse missingness behaviors. Quantitatively, SRP reduces an adversarial discriminator's detection AUROC by 4\% compared to the baseline, indicating that the synthesized relational structures and temporal patterns are significantly harder to differentiate from real-world data. Finally, when used to pretrain Relational Transformers, SRP consistently improves downstream zero-shot classification and regression performance, demonstrating its superior utility for scalable representation learning.