Support constraints for the PAR Synthesizer

[schweima]

Environment:
Python 3.7.10
sdv 0.11

Error description:
There is no documentation detailing whether PAR supports constraints.

Is there a way similar to the HMA1 approach mentioned in #296 ?

Initial tests with according to this workaround referenced by @npatki in #566 with the Between constraint are either ignored or yield NaN values when sampling.

Additional information:
Any hints appreciated. Thank you! Use case: measurement time series with a lot of columns that have to be >= 0.

[sokol11]

Hi. I am also trying to use constraints with PAR. I need to have the UniqueCombinations constraint in my case. Would you mind sharing the code you used to try the HMA1 workaround? It is giving me a KeyError when I naively try:

from sdv.timeseries import PAR
from sdv.constraints import UniqueCombinations
from sdv import Metadata

constraints = [UniqueCombinations(columns=['date', 'col1'])]
metadata = Metadata()
metadata.add_table(name='data')
metadata._metadata['tables']['data']['constraints'] = constraints
model = PAR(metadata, entity_columns=['col1'], sequence_index='date', epochs=128, cuda=True)

I get:

model.fit(data)

...

KeyError: Metadata
  root_path: .
  tables: ['data']
  relationships:

Thank you!

[npatki]

The PARModel does not currently support constraints, so any workaround/code that was used for HMA1 will not work here.

Use case: measurement time series with a lot of columns that have to be >= 0.

Issue #745 is specifically tracking the feature request for supporting min/max bounds on the PARModel.

However, let's continue to keep issue open for tracking the feature request of adding general constraints on sequential data.

[tosador]

Hi, I am also trying to use the PAR model setting inequality constraints on sdv==1.0.

Inequality constraints are not supported yet (#1320).

I noticed that if I try to generate financial OHLC bars data, sometimes the constraints:

• H >= C >= L
• H >= O >= L

are not satisfied.

Use case: training/backtesting machine learning models on synthetic financial bars data.

[Ng-ms]

Hello
In my dataset, i am having certain columns should be generated together, for example, the beginning of the treatment the end of the treatment and the code of the treatment, and the name of the treatment should be predicted together, now the PARSynthesizer is generating them all separate and does not follow this logic in the real data, the only thing I can think of it to solve this is to have constraints in the synthesizer to deal with these columns as ´package ´ to generate them together

[Ng-ms]

is there any updates on this request?

[npatki]

Hi @Ng-ms, thanks for describing your use case. In theory, this type of logic would be well-served by the FixedCombinations constraint.

I am curious if for any of the columns you mention (beginning of treatment, end of treatment, code of treatment) are you marking any as context_columns in PAR?

There seems to be demand for this feature. I think it will be easier to accommodate constraints in PAR if either all of the columns are context columns or if all of the columns are NOT context columns. I am interested to know if that would cover most of the cases that folks have encountered?

[Ng-ms]

in this use case, these columns are not context columns.

[srinify]

Hi @schweima @sokol11 @tosador @Ng-ms we've actually added partial support for constraints in PARSynthesizer in SDV 1.14.0: https://github.com/sdv-dev/SDV/releases/tag/v1.14.0

You can use PARSynthesizer with constraints as long as the columns involved with the constraint are either ALL context columns or ALL non-context columns. We've also added a section in our docs for this under the PAR Synthesizer specific section: https://docs.sdv.dev/sdv/sequential-data/modeling/customizations

We're hoping this helps a few of you -- let us know!