Memorization#
These metrics compare generated records with the actual data used to train the generator. They measure different aspects of proximity:
Metric |
Query direction |
Question |
|---|---|---|
Generated and test → training |
Are generated records closer to training than independent real records? |
|
Generated → training |
What fraction of generated rows are at a small distance from the training set? |
|
Training → generated |
What fraction of training rows are at a small distance from the generated set? |
All three use exact distances and return point-only results. DCR ratio requires
independent test data for its denominator and includes distance percentiles.
The two rates need only train and generated; their geometry and values
depend on these two pools.
Near-match and replication use the same fixed inclusive Euclidean threshold in the training-fitted distance space. Repeated generated copies each contribute to near-match rate, while a training row counts only once toward replication ratio. Dense or discrete real populations can naturally produce close records, so these measurements need population context.