Ensure values fall within the specified range

[dyuliu]

This is a follow-up question of the following issue:
#200 (comment)

My situation:
I have many columns:

• they are non-negative
• they have a lot of 0 value
• they sometimes have large positive values

My synthetic model often generates negative values for these columns.

My questions:
(1) I am not able to replicate by using CopulaGAN with "gamma" distribution to control the value range. (#200 (comment))

(2) two solutions here are difficult for me.
The first solution is super slow because I have too many columns having this issue.
The second solution is not clear to me, @csala, can you can me the detailed example?

[MLjungg]

I have also experienced problems with setting "gamma" manually to ensure the distribution falls in the positive range. An alternative is to set the distribution to "semi_bounded", which allows other distributions to be selected that are defined in the positive range – that has worked for me.

[npatki]

I'd also recommend learning different distributions if possible. That will ensure your correlations remain undisturbed.

(2) two solutions here are difficult for me.
The second solution is not clear to me, @csala, can you can me the detailed example?

In the second solution, you will write a CustomConstraint will strategically transform and/or reverse_transform the data to fit your constraints.

One possible way to do this would be to "fix" any data that doesn't conform to the constraint.

# retroactively fix the output data to conform to your constraint
def reverse_transform(self, table_data):
    # every value in Col A that is less than 0 gets fixed to be 0; leave rest of the data alone
    table_data.loc[table_data['col_A'] < 0, 'col_A']  = 0

I'm not exactly sure what Carles' suggestion was, but I assume it's a similar strategy where by either transforming the input data and/or fixing the output data, you will ensure constraints are met.

The issue with any of these approaches is that you're affecting any correlations by using non-linear transformations.

[npatki]

Closed in #492 -- by default, the models will respect the min and max values observed in the real data, although you have the ability to reset this as needed.