Synthetic Data Vault (SDV)

SDV handles heterogeneous tabular data containing a mix of continuous, discrete, and categorical columns with complex, non-Gaussian distributions. It uses hierarchical generative models to learn column-specific marginal distributions and a copula-based dependency structure to capture inter-column correlations.

• Models multimodal continuous columns (e.g., distributions with multiple peaks)
• Handles severe class imbalance in categorical variables
• Preserves semantic constraints and business logic (e.g., end_date > start_date)
• Uses CTGAN for conditional generation and GaussianCopula for statistical modeling

SDV extends generation beyond flat files to relational databases with parent-child dependencies. The Hierarchical Modeling Algorithm (HMA) recursively traverses table relationships, learning to generate rows in child tables that maintain referential integrity with synthetic parent records.

• Preserves primary key-foreign key relationships across tables
• Models one-to-many and many-to-many cardinalities
• Generates consistent synthetic databases where joins remain valid
• Handles star schemas and normalized transactional structures

SDV captures temporal dynamics and sequential dependencies in time-series data. It distinguishes between static attributes (fixed per entity) and time-varying attributes (changing over sequence steps), generating realistic trajectories that preserve autocorrelation patterns.

• Models irregular sampling intervals and variable-length sequences
• Preserves trend, seasonality, and cyclical patterns
• Handles entity-specific temporal behaviors
• Supports conditional generation based on static attributes

SDV includes a comprehensive evaluation framework that quantifies synthetic data across three critical dimensions: statistical fidelity, downstream utility, and privacy protection. This enables data custodians to make evidence-based decisions about data release.

• Statistical similarity: Compares column distributions and pairwise correlations
• Utility metrics: Train-Synthetic-Test-Real (TSTR) and feature importance similarity
• Privacy metrics: Nearest Neighbor Adversarial Accuracy (NNAA) and membership inference risk
• Generates quality reports with visualizations and diagnostic scores

SDV uses a declarative metadata specification that describes the data's structure, column types, constraints, and relationships. This metadata acts as a configuration layer, enabling the system to automatically select appropriate models and enforce data validity rules during generation.

• Auto-detects primary keys, foreign keys, and alternate keys
• Enforces regex patterns, value ranges, and categorical sets
• Supports anonymization rules for PII (personally identifiable information)
• JSON-based metadata format enables version control and auditability

SDV provides a library of pre-trained generative models optimized for specific data types and structures. These synthesizers are built on state-of-the-art architectures including GANs, variational autoencoders, and copula-based methods, each selected based on the metadata characteristics.

• CTGAN: Conditional GAN for complex tabular data with mixed types
• TVAE: Variational autoencoder for tabular data
• GaussianCopula: Fast statistical model for simpler distributions
• PAR: Probabilistic autoregressive model for sequential data

The foundational Conditional Tabular GAN within the SDV ecosystem. It uses a mode-specific normalization technique to handle non-Gaussian, multimodal continuous columns and a training-by-sampling strategy to address class imbalance in discrete columns. This allows it to faithfully model the complex, heterogeneous distributions found in real-world enterprise tables, such as financial transactions or electronic health records.

The Tabular Variational Autoencoder is an alternative generative model in the SDV suite. Unlike GANs, it learns a probabilistic latent representation of the data. It is particularly effective for datasets where GAN training suffers from mode collapse. TVAE uses a beta-VAE objective to balance reconstruction fidelity and latent space regularization, making it a robust choice for generating high-dimensional tabular data with complex inter-column correlations.

A Probabilistic AutoRegressive model designed for sequential and time-series data within SDV. It models each column as a conditional distribution given previous columns and a context vector. This architecture explicitly captures temporal dynamics and causal ordering, making it suitable for generating synthetic multi-session user logs, sensor telemetry, or patient treatment histories where the sequence of events is critical.

The companion library for evaluating synthetic data quality. It provides a comprehensive suite of metrics across three pillars: Statistical Fidelity (e.g., Column Shapes, Correlation Similarity), Utility (e.g., ML Efficacy via TSTR), and Privacy (e.g., Nearest Neighbor Distance Ratio). It generates a diagnostic report that quantifies the trade-off between realism and re-identification risk, enabling data scientists to validate their synthetic data before downstream use.

A critical abstraction layer that defines the relational structure and semantic constraints of a dataset before modeling. Users specify primary keys, foreign keys, column data types (e.g., categorical, datetime), and business logic. This metadata object guides the generative process, ensuring that synthetic data respects referential integrity and domain-specific rules, such as a patient's discharge date always occurring after their admission date.

The Hierarchical Modeling Algorithm is the core engine for generating synthetic multi-table relational databases. It recursively traverses the database schema defined in the metadata, modeling parent tables first and then conditioning child table generation on the synthetic parent rows. This preserves referential integrity and cross-table correlations, allowing the creation of a fully synthetic, structurally identical replica of a complex enterprise database.