Interpretation of 'nearest synthetic neighbor distance' metric about synthcity

[bcebere]

Thank you for reporting this.

Indeed, the docs for "nearest_syn_nigh_distance" are confusing.

The 'common rows proportion' metric is a sanity check, and it tests if any of the training items are identical in the sampled data.

The nearest_syn_nigh_distance trains a KNeighbors model in the synthetic data and reports the closest distance for the training data. These distances are further normalized and then averaged, resulting in the score. If the mean is 0, then all the training data are replicated in the synthetic data. At the same time, if the distances are 1(normalized), it means all the synthetic samples are very far away from the training data. This is not good either.

We will try to clarify this metric and improve the docs.

Thanks!