Principle:Sdv dev SDV Single Table Quality Evaluation
| Knowledge Sources | |
|---|---|
| Domains | Data_Quality, Statistics, Evaluation |
| Last Updated | 2026-02-14 00:00 GMT |
Overview
A statistical evaluation framework that measures how faithfully synthetic data reproduces the distributions and correlations of real single-table data.
Description
Single-table quality evaluation compares real and synthetic data across multiple statistical dimensions: column-level marginal distributions (Column Shapes) and pairwise column correlations (Column Pair Trends). The framework produces a QualityReport with an overall score between 0 and 1, where 1 indicates perfect statistical fidelity. Individual column and column-pair scores can be inspected for detailed analysis.
The evaluation is built on the SDMetrics library, which provides standardized statistical tests for synthetic data quality.
Usage
Use quality evaluation after generating synthetic data to assess how well it preserves the statistical properties of the real data. This is the standard final step in any SDV synthesis workflow.
Theoretical Basis
The quality score aggregates two metric categories:
Column Shapes: Per-column distributional similarity, measured using KSComplement (numerical/datetime) or TVComplement (categorical/boolean).
Column Pair Trends: Pairwise correlation similarity, measured using CorrelationSimilarity (numerical pairs) or ContingencySimilarity (categorical pairs).
Overall score = average of Column Shapes score and Column Pair Trends score.