Synthetic Data Vault (SDV)

SDV provides specialized models for different data types, each designed to handle unique structural challenges:

• CTGAN: A conditional GAN that models non-Gaussian, multi-modal distributions in tabular data with mixed discrete and continuous columns
• TVAE: A variational autoencoder adapted for tabular data, learning a probabilistic latent representation to generate novel rows
• PAR: A probabilistic autoregressive model for synthesizing multi-sequence time-series data while preserving temporal dynamics
• HMA1: A hierarchical modeling algorithm that captures relationships across relational databases with multiple interconnected tables

SDV's Hierarchical Modeling Algorithm (HMA) synthesizes entire relational databases while maintaining foreign key constraints and referential integrity:

• Models parent tables first, then conditions child table synthesis on generated parent keys
• Preserves primary key-foreign key relationships across all tables
• Ensures synthetic child records reference valid synthetic parent records
• Handles complex star schemas and snowflake schemas common in enterprise data warehouses
• Prevents orphaned records that would break downstream analytics

SDV integrates with SDMetrics to provide comprehensive quality reports that quantify how well synthetic data mirrors real data:

• Column Shapes: Measures how closely univariate distributions match using KSComplement and TVComplement metrics
• Column Pair Trends: Evaluates bivariate correlations and associations using correlation similarity scores
• Boundary Adherence: Checks that synthetic values respect the min-max ranges and domain constraints of real data
• Statistical Similarity Score: An aggregate metric combining all dimensions into a single 0-1 fidelity score

SDV enforces business rules and logical constraints during synthesis to ensure generated data is operationally valid:

• Unique constraints: Guarantees no duplicate values in specified columns
• Inequality constraints: Enforces relationships like column_a > column_b
• Range constraints: Bounds numeric values to valid operational ranges
• Custom logic: Supports user-defined Python functions for complex validation rules
• Constraints are applied during both model fitting and data generation phases

SDV uses a declarative metadata JSON specification to describe dataset structure before modeling:

• Defines column data types, primary keys, and foreign key relationships
• Specifies semantic types (e.g., email, address, datetime) for domain-aware synthesis
• Supports datetime formats, anonymization policies, and column-level privacy settings
• Enables reproducible synthesis pipelines by version-controlling metadata alongside code
• The metadata object serves as the single source of truth for the entire synthesis workflow

The Conditional Tabular GAN is the foundational neural network within the SDV ecosystem designed specifically for non-Gaussian, multi-modal tabular data. It introduces a training-by-sampling technique to handle class imbalance and a mode-specific normalization to overcome the limitations of min-max scaling on complex columns. CTGAN models mixed discrete and continuous columns effectively, making it the default engine for synthesizing single-table relational databases.

The official open-source Python library for evaluating synthetic data quality. It provides a standardized reporting framework that quantifies three pillars:

• Statistical Fidelity: Measures column shapes and pair trends against real data.
• Privacy Risk: Detects membership inference and re-identification vulnerabilities.
• Utility: Assesses downstream machine learning performance using the Train-Synthetic-Test-Real (TSTR) paradigm.