UniqueCombination constraint with numerical values

[jrmelog]

Environment details

• SDV version: 0.9.1
• Python version: 3.8.2
• Operating System: Windows

Problem description

Is it possible to use the UniqueCombination constraint with numerical columns on CTGAN? I tried to use it, but it shows an error: Can only use .str accessor with string values! when I try to fit the model.

What I already tried

I specified the fields types:

field_types = { "Duracao do Acordo Informada": {"type": "numerical", "subtype": "float"}, "Duracao do Acordo Efetiva": {"type": "numerical", "subtype": "float"}, "Situacao do Acordo": {"type": "numerical", "subtype": "float"} }

and then created the constraint:

acordo_constraints = UniqueCombinations(columns=['Duracao do Acordo Informada', 'Duracao do Acordo Efetiva', 'Situacao do Acordo'], handling_strategy='transform')

constraints = [acordo_constraints]

and created the model:

model = CTGAN(epochs=100, field_types = field_types, constraints = constraints)

[npatki]

I'm able to reproduce using this test set. This limitation isn't mentioned in the docs, so I'm not sure if this is a bug in the code or docs. If it's the latter, we can turn this into a feature request.

My workaround was to recast the columns as strings, so they become categorical.

constraint = UniqueCombinations(
    columns=['name', 'age'],
    handling_strategy='transform')

model = GaussianCopula(constraints=[constraint])

# need to recast before fitting the model
table['age'] = table['age'].astype(str)

model.fit(table)

Out of curiosity, could you provide more details about your use case?

I can think many scenarios where categorical strings are UniqueCombinations, like country/city pairs. However, I can't think of any where float values need to uniquely appear. Floats can have many decimal points, so it seems rare to me that a value like 1.23456789 must appear precisely & uniquely with another value.

1. Are you float values rounded & bounded? For eg, they only appear as full integers from 1.0-10.0
2. Do your float values actually encode categorical entities such as a status?

[csala]

so I'm not sure if this is a bug in the code or docs. If it's the latter, we can turn this into a feature request.

I would consider this a feature request. It was already requested on #196 too, btw, so if we agree that we are talking about the same I would flag this one as a duplicate and close, in favor of the other one.

However, I can't think of any where float values need to uniquely appear.

There could be a scenario in which there are unique combinations of products and prices. For example, think about a table that contains hotel reservations that include the columns room_type and price_per_night. One may want each room_type to always show up with the same price_per_night, which should be seen as a given "constant".

[npatki]

Agreed. Closing out this issue in favor of the older feature request.

@jrmelog, If would be great if you could provide more detail about your use case in #196. That way, we can make sure we factor in whatever nuances we find when we provide a fix.