Add LogScaler transformer

[frances-h]

Problem Description

Currently, the ScalarRange, ScalarInequality, Positive, and Negative constraints all scale the data using log-based transformers. Since the SDV already enforces min/max values by simpler means, these constraints can be deprecated. However, it would be helpful to move the scaling logic into the RDT library, so that it may be used independently of the constraints in the future.

Expected behavior

Create a new RDT called LogScaler. This RDT transforms the data by applying a log. Its functionality is equivalent to the ScalarInequality constraint (as well as Positive and Negative).

Parameters

• missing_value_replacement (object): Same as FloatFormatter
• missing_value_generation (str or None): Same as FloatFormatter
• constant (float): The constant to set as the 0-value for the log-based transform. Default to 0 (do not modify the 0-value of the data)
• invert (bool): Whether to invert the data with respect to the constant value. If False, do not invert the data (all values will be greater than the constant value). If True, invert the data (all the values will be less than the constant value). Defaults to False.
• learn_rounding_scheme (bool): Same as FloatFormatter

Implementation Notes

• Apply similar logic that is used in ScalarInequality
• During fit:
  • Learn needed information about missing values and rounding scheme.
  • Validate that we can actually take the log of the data.
• During transform:
  • Fill in missing values
  • Transform the data
    • If invert is False (default): transformed_data = log(data - constant)
    • If invert is True: transformed_data = log(constant - data)
  • If there are any issues with taking the log, raise a descriptive error explaining what wen wrong:
    • Example: Error: Unable to apply a log transform to column 'capital-gains' due to a non-positive value (-1).

Intended usage: A user would call update_transformers to replace the FloatFormatter with the LogScaler for any columns that would benefit from this.

synthesizer = GaussianCopulaSynthesizer(metadata)
synthesizer.auto_assign_transformers(data)
synthesizer.update_transformers({
  'capital-gains': LogScaler(missing_value_replacement=0, constant=-1e-5)
})
synthesizer.fit(data)