Gaussian Copula is generating different data with metadata and without metadata.

[kvrameshreddy]

When we create a Gaussian copula model without defining metadata explicitly, the generated synthetic data holds the properties of the sample data( like min and max values). When we define the same by explicitly passing meta data the generated data is loosing the properties.

from sdv.tabular import GaussianCopula

df1=pd.read_csv("rounding_data_test.csv")
meta_default={'fields': {'COL1': {'type': 'numerical',
                                  'subtype': 'float'} ,
                         'COL2': {'type': 'numerical',
                                   'subtype': 'float'},
                          'COL3': {'type': 'numerical',
                                    'subtype': 'float'},
                          'COL4': {'type': 'numerical',
                                    'subtype': 'float'}}}
model1 = GaussianCopula(table_metadata=meta_default)
model1.fit(df1)
new_data_with_meta = model1.sample(100)

''' without meta data'''

model2 = GaussianCopula()
model2.fit(df1)

new_data_without_meta = model2.sample(100)

rounding_data_test.csv

From the above screenshot consider the 'col2' column in the original data(df1) min value is 8.37 and max is 196.8 .
now the generated new_data_without_meta holds the min and max values of the sample but the new_data_with_meta generates negative values. I even verified for the distributions of the data for each column, both the models have same distributions.

The same behaviour is seen with constraints also, when passed directly gives similar data, when passed with explicit metadata gives out of bound values.
can you please look into this issue.
Thankyou.

[katxiao]

Hi @kvrameshreddy, thanks for your question!

I agree that we should expect min and max values to be automatically detected the same way in both cases (when metadata is provided and when it is not). We will look into this issue soon.

[kvrameshreddy]

Hi @katxiao , thanks for looking into this issue, When can we expect 0.12.1 release ?

Thankyou

[katxiao]

Hi @kvrameshreddy, it will go out this week!