Declare columns as value-safe

[yoid2000]

Often the values of a column are publicly known, and there is no reason to suppress them.

I want to be able to tag these columns as such in the API, and then declare them value-safe in the microdata class.

I want this to work for all column types. A side-effect of this is that, if a numeric or datetime column is declared value-safe, then only those values will be used in microdata creation. I'm tempted to implement this by using the numeric for sorting, then converting to string. This is a bit weird. My concern is that, when converting to string, any sense of numeric value is lost---a value might be pushed to a very far-away value simply because there is a gap in the continuous sequence.

Another approach for continuous data would be to build continuous microdata, but then afterwards force each value to the nearest safe value. A possible problem here is that we could have several outliers in a given range, and we'd need to pick one. But this doesn't seem much of a problem ... since assignment in a range is random, then effectively we'd be picking a random value, more or less. This sounds like the right behavior.

[yoid2000]

This means that I need to have a clean way to convey this and potentially other information about columns in the API. An example would be the ability to declare two columns as latitude/longitude.

How about:

sdx = Synthesizer(df_orig, value_safe_columns=['col1', 'col2', ...])

Seems good.

[yoid2000]

Ok, let's rethink this.

The normalized values must not be modified (rounded) because these are needed to return the original values from the normalized values.

In any event, we will map to the closest normalized value when retrieving safe values, so we will get the exact normalized value, which will return very nearly the exact original value. Then we'll do rounding, so any machine error in the reverse-normalized value will be removed.