Most research in synthetic data generation treats data as a single, flat table. However, real-world information—from electronic health records to financial ledgers—rarely lives in isolation. It resides in relational databases: complex webs of tables interconnected by primary and foreign keys.
When we move from synthesizing a single table to an entire relational schema, the challenge shifts from merely matching a statistical distribution to maintaining referential integrity. If a synthetic "Transaction" table references a "Customer ID" that does not exist in the synthetic "Customer" table, the database becomes logically broken and useless for downstream tasks like SQL querying or software testing.
The Challenge of Multi-Hop Dependencies
In a relational database, the most important correlations often span across multiple tables. For example, a customer’s age (Table A) might influence their likelihood of purchasing a specific category of product (Table C), even if those tables are only linked through an intermediate "Orders" table (Table B).
Masoumeh Shafieinejad and Xi He identify these as multi-hop dependencies. Traditional synthesis methods that process tables one by one often fail to capture these long-range interactions. As noted in ClavaDDPM: Multi-relational Data Synthesis with Cluster-guided Diffusion Models, ignoring these correlations results in a synthetic dataset that looks correct locally (within each table) but lacks global coherence.
From Sequential to Holistic Generation
To solve the integrity problem, researchers have moved away from independent sampling toward structured, dependency-aware pipelines.
- Sequential Generation: Early frameworks like PrivBayes, developed by Jun Zhang and Xiaokui Xiao, use Bayesian networks to model attribute correlations. In a multi-table context, this often involves a "parent-first" approach: generating the primary entities first and then conditioning the generation of "child" records on those existing parents.
- Graph-Based Ordering: Frameworks like LAKEGEN use graph-based algorithms to determine the exact order of synthesis. By following the direction of foreign key constraints, the model ensures that every foreign key value generated has a corresponding "anchor" in a previously synthesized primary key column.
- Simultaneous Synthesis: To avoid the "error propagation" inherent in sequential steps, newer models like Graph Conditional Flow Matching for Relational Data Generation attempt to generate the entire database content across all tables simultaneously. This holistic approach allows the model to respect complex "diamond" relationships (where two tables are linked through multiple paths) that are often lost in ordered generation.
Neurosymbolic and LLM-Based Approaches
A recent shift involves using Large Language Models (LLMs) and neurosymbolic frameworks to bridge the gap between unstructured intent and structured data.
Mushtari Sadia and Ang Chen introduced SQUiD, which synthesizes full relational databases from unstructured text. By decomposing the task into modular stages—schema generation, entity extraction, and relationship assignment—they ensure that even smaller models can maintain referential integrity. Similarly, Milad Abdollahzadeh demonstrated that instruction-tuning compact LLMs on high-quality metadata can overcome the "structural mismatch" between 1D text-based models and 2D relational data.
Handling Complex Hierarchies
Real-world data is rarely a clean tree of one-to-many relationships. Many-to-many relationships (e.g., students enrolled in many classes, and classes containing many students) present a significant hurdle for standard synthesizers.
The DEREC-SIMPRO framework, proposed by Tung Sum Thomas Kwok and Chi-Hua Wang, addresses this by algorithmically restructuring many-to-many datasets into a one-to-many hierarchy. By creating an artificial "parent" table that contains shared contextual attributes, they allow standard multi-table synthesizers to process complex relational data that lacks a natural hierarchy.
Privacy and Utility in Relational Space
Finally, synthesizing relational data is not just about structure; it is about safety. PrivBayes pioneered the use of Differential Privacy (DP) in this space, injecting noise into low-dimensional marginals to protect individual records.
However, as Chaoyi Zhu notes in the context of DP-TLDM, applying DP to relational databases is significantly harder than to single tables. The model must protect the privacy of individuals whose information may be scattered across dozens of linked tables, requiring a delicate balance between adding enough noise to ensure privacy and maintaining enough signal to preserve the intricate joins that make relational data valuable.