Unwanted negative numbers generated

[wbarnsley]

• SDV version: 0.2.2
• Python version: 3.7.3
• Operating System: mac

Description

I was trying to generate a synthetic version of a table. The table in question has two numeric columns, of which all of the entries are positive. After fitting the sdv object, I sampled and found that the two numeric columns contained negative values. After reading your paper from IEEE I was under the impression a truncated gaussian would be used and that this should not happen.

What I Did

import pandas as pd
from sdv import SDV, Metadata

df_data = pd.DataFrame(data = {'col_1':list(range(10))})

sdv = SDV()
table = {'data': df_data}
metadata = Metadata()
metadata.add_table('data', data=table['data'])
sdv.fit(metadata, table)
df_synth = sdv.sample_all(len(df_data.index))['data']

print(df_synth)

Output

   col_1
0     10
1     -3
2      0
3      8
4      3
5      9
6      6
7      8
8      4
9      7

[csala]

Hi @wbarnsley the truncated gaussian distribution was indeed not being used before.
We are currently doing some changes here and on the Copulas library that will make using the Truncated Gaussian distribution on SDV.
We will let you know here once the changes are all in and ready to use.
Stay tuned!

[csala]

This topic is being covered in a newer issue, so I'm closing this to follow up in one place only.