Numerical column considered as regional?

[miaoli-04]

Environment Details

Please indicate the following details about the environment in which you found the bug:

• CTGAN version: 1.25.0 (sdv)
• Python version: 3.12.8
• Operating System:

Error Description

When trying the fit the "automobile" dataset from UCIML, the 'city-mpg' column, which is continuous, seems to be interpreted as a location and a column of strings is generated in the synthetic data. This might have to do with the column name, as if I rename the column as 'mpg', column of correct datatype will be returned.

Steps to reproduce

import pandas as pd
from sdv.single_table import CTGANSynthesizer
from sdv.metadata import Metadata

from ucimlrepo import fetch_ucirepo 
  
# fetch dataset 
automobile = fetch_ucirepo(id=10) 
  
# data (as pandas dataframes) 
X = automobile.data.features 


metadata = Metadata.detect_from_dataframe(X)

synthesizer = CTGANSynthesizer(metadata)
synthesizer.fit(X)
synthetic_data = synthesizer.sample(num_rows=1000)

[npatki]

Hi @miaoli-04 thanks for filing this issue and providing the details. The root cause is that metadata is being incorrectly auto-detected in the following line.

metadata = Metadata.detect_from_dataframe(X)

Note that CTGANSynthesizer is correctly creating synthetic data based on this metadata. So this does not appear to be a bug in the CTGAN algorithm/library itself.

Recommended Approach

SDV's metadata auto-detection algorithm makes some judgements based on column names and the data. The metadata auto-detection is not guaranteed to be accurate or complete, so we always recommend inspecting the metadata and updating it to accurately reflect your data. (Then you can save the corrected metadata so that you won't have to go through this process again for the dataset.)

The recommended approach is to update metadata to accurately reflect the dataset. You can refer to this section of our SDV docs to understand more about inspecting and updating metadata. Additionally, this guide on sdtypes might be useful.

metadata = Metadata.detect_from_dataframe(X)

# inspect and update any inaccuracies in the metadata
metadata.update_column(
  column_name='city-mpg',
  sdtype='numerical')

# save the metadata so you won't have to do this again
metadata.save_to_json('automobile_metadata.json')

# now use it when using CTGAN
synthesizer = CTGANSynthesizer(metadata)
...

CTGANSynthesizer uses the metadata to know which columns to learn patterns from (numerical and categorical), and which columns to anonymize. So it's critical that the metadata is accurate.

Enhancing the metadata auto-detection

SDV's metadata auto-detection feature can definitely be enhanced. From your example, I can see that certain sdtypes (such as "city") should not be assigned if the data is present as a int/float. I will file a feature request based on the issue you are observing so that we can get this resolved in the future. Thanks!