Synthetic Data Vault (SDV)

A core model within the SDV ecosystem specifically designed for tabular data with mixed discrete and continuous columns. CTGAN uses a conditional generator and training-by-sampling to overcome the class imbalance typical in real-world datasets. It applies a mode-specific normalization technique to handle non-Gaussian, multimodal continuous distributions, ensuring the synthetic data faithfully preserves column-wise statistical properties and correlations.

A synthesizer for multi-table, relational datasets that models parent-child relationships without flattening the data. PAR learns the distribution of each table and the referential integrity between them. It generates synthetic data by first sampling parent tables and then conditionally generating child table rows, ensuring that foreign key relationships and relational structures are statistically preserved in the synthetic output.

SDV provides specialized models for sequential and time-series data, such as the DoppelGANger and PAR architectures. These models capture complex temporal dependencies, including autocorrelation, seasonality, and trends. They are designed to handle the unique challenges of time-series generation, such as variable sequence lengths and the need to preserve the conditional distribution of a point given its historical context.

SDV includes a built-in evaluation framework to quantitatively measure the quality of generated data. It produces a Quality Report that compares the synthetic data to the real data across three pillars:

• Column Shapes: Statistical similarity of individual column distributions.
• Column Pair Trends: Preservation of bivariate correlations.
• New Row Synthesis: A privacy check ensuring rows are novel and not direct copies. This framework provides a standardized score for data utility.

SDV allows users to define logical constraints on the synthetic data generation process to enforce business rules and enhance privacy. Constraints like UniqueCombinations, GreaterThan, and FixedIncrements ensure that the synthetic data adheres to the same domain logic as the real data. This prevents the generation of nonsensical or impossible records, which is critical for maintaining utility in downstream machine learning tasks.

The entire SDV ecosystem is orchestrated by a metadata JSON specification that describes the structure, data types, and relationships of the source database. This metadata acts as a blueprint, allowing the synthesizers to automatically understand which columns are primary keys, which are categorical, and how tables are connected. This design enables a programmatic and reproducible data generation pipeline without manual feature engineering.