For the complete documentation index, see llms.txt. This page is also available as Markdown.

BoundaryAdherence

This metrics measures whether a synthetic column respects the minimum and maximum values of the real column. It returns the percentage of synthetic rows that adhere to the real boundaries.

Data Compatibility

  • Numerical : This metric is meant for numerical data

  • Datetime : This metric converts datetime values into numerical values

If you have missing values in the real data, then the metric will consider them valid in the synthetic data. Otherwise, they will be marked as out-of-bounds.

Score

(best) 1.0: All values in the synthetic data respect the min/max boundaries of the data

(worst) 0.0: No value in the synthetic data respects the min/max boundaries of the data data

The graph below shows an example of some fictional real and synthetic data (black and green, respectively) with BoundaryAdherence=0.912.

The real data is in range [37.0, 97.7], shown by vertical dotted lines. Around 8.8% of the synthetic data is outside of the min/max bounds of the real data, so the overall BoundaryAdherence score is 91.2%.

How does it work?

This metric computes the min and max values of the real column. Then, it computes the frequency of synthetic values that are in the [min, max] range. If there are missing values in the real data, then the metric will consider them as valid in the synthetic data; otherwise, missing values are considered out-of-bounds.

In some cases the allowable min/max range as well as the allowable missing values may be more expansive than what is present in the real data. This can happen when your real data is actually a smaller subset of the entire database. In this case you can override the allowable ranges by providing the range_min, range_max, and range_is_nullable parameters. If provided, these values will be used for checking the synthetic data rather than ranges derived from the real data.

Usage

To manually apply this metric, access the single_column module and use the compute method.

Parameters

  • (required) real_data: A pandas.Series object with the column of real data

  • (required) synthetic_data: A pandas.Series object with the column of synthetic data

  • range_min: A float or datetime containing the absolute minimum value that the data can possibly have. This may be less than the minimum value present in the real training data. If provided, this value will be used instead of the minimum value from the real_data.

  • range_max: A float or datetime containing the absolute maximum value that the data can possible have. This may be greater than the maximum value present in the real training data. If provided, this value will be used instead of the maximum value from the real_data.

  • range_is_nullable: A boolean describing whether the column is allowed to be null. This can be True even if the real training data does not have any missing values. If provided, this value will be used instead of computing it from the real_data.

FAQs

Is there an equivalent metric for discrete data?

For discrete datasets, only a select number of values are possible. Use the CategoryAdherence metric to ensure they have the correct values.

What kind of scores should I expect to see?

The scores that you see are based on the model that you used to create the synthetic data. Some models are designed to learn the min/max boundaries from the real data. For such models, the score should always be 1.0

Other models do not learn the boundaries. In these cases, you can expect the score to be close to, but not exactly, 1.0. Note that this may be a desirable feature if it is risky to leak the true min/max values in the synthetic data.

What if my synthetic data is only covering a small subset of the real values?

This metric only quantifies cases where the synthetic data is going out of bounds. If you're interested in knowing whether the synthetic data covers the full range of real values, use the RangeCoverage metric.

Last updated