Metrics Glossary

Each metric in this section quantifies a different aspect of the synthetic data.

Browse by use case

We recommend applying the metrics using SDMetrics Reports. Reports determine which exact metrics to use based on your data.

Diagnostic

These metrics capture basic, statistical properties of synthetic data that we expect to be captured.

Get started using the Diagnostic Report.

Or apply these metrics individually:

• BoundaryAdherence, CategoryAdherence: measure the validity of statistical values

• KeyUniqueness: measure the validity of primary keys

• ReferentialIntegrity, CardinalityBoundaryAdherence: measure the validity of a connection between a foreign and primary key

• TableStructure: measure whether the overall structure of the data is the same

• MissingValueSimilarity, StatisticSimilarity: compare data using basic statistics

• CategoryCoverage, RangeCoverage: measure whether the overall synthetic data spans the possibilities

• SequenceLengthSimilarity, StatisticMSAS: compares the quality of real and synthetic data that represents sequential information

Quality

These metrics capture the overall patterns and shapes of your data. While these may not always be 100%, a higher score generally indicates higher quality data.

Get starting using the Quality Report.

Or apply these metrics individually:

• KSComplement, TVComplement: compare shapes (aka marginal distributions, histograms)

• ContingencySimilarity, CorrelationSimilarity: compare 2D distributions & pairwise correlations

• (Multi table only) CardinalityShapeSimilarity: compare parent/child connections

Privacy

These metrics capture the privacy of your data. Use them to explore privacy protection and risks associated with different types of attack scenarios.

• DisclosureProtection, DisclosureProtectionEstimate: Measure the risk associated with broadly sharing your synthetic data

ML Augmentation

These metrics capture the value of using synthetic data for the purposes of training an ML model. They assume you are augmenting the real data with synthetic data together to create a more enhanced training set.

• BinaryClassifierPrecisionEfficacy: Use agumented data to train a binary classifier, optimized for precision

• BinaryClassifierRecallEfficacy: Use augmented data to train a binary classifier, optimized for recall

Browse by data granularity

Different metrics compare data at various levels of granularity.

Single Column

These metrics compare single columns of data.

• CategoryCoverage, RangeCoverage: measure data coverage

• BoundaryAdherence, CategoryAdherence: measure whether the synthetic data is chosen from valid boundaries or lists of options

• KSComplement, TVComplement: compare shapes (aka marginal distributions, histograms)

• StatisticSimilarity, MissingValueSimilarity: compare data statistics

• KeyUniqueness: measures whether the primary keys are unique

These metrics compare single columns that represent sequential information.

• SequenceLengthSimilarity: compares the length of sequences between real and synthetic data

Column Pairs

These metrics compare pairs of columns.

• ContingencySimilarity: compare 2D distributions

• CorrelationSimilarity: compare pairwise correlations

• ReferentialIntegrity: compute the validity of a connection between a foreign key and primary key

These metrics compare pairs of columns that represent sequential information.

• StatisticMSAS: compare summary statistics between sequences of real and synthetic data

Single Table

These metrics compare a table of data.

• DisclosureProtection, DisclosureProtectionEstimate: Measure the risk associated with broadly sharing your synthetic data

Multi Table

These metrics compare multiple tables of data.

• CardinalityShapeSimilarity: compare parent/child connections

More Metrics

Can't find what you're looking for? Check out the Metrics in Beta for additional options.