Documentation / standardization for BaseFeaturizer about matminer

[computron]

So to follow up on this:

For the documentation, I don't have much guidance at this time. If someone has some proposals of what they'd like to see, I can provide some feedback.

As for nan vs exception, I think the behavior I'd most like is for an individual featurize() function (for one object) to throw an exception if something goes wrong. For example, if the featurizer requires cations and anions (needs + and - charged sites) but you feed in a metal where all the charges are zero - to throw a meaningful exception like RuntimeError("no cation site found"). Hopefully not just a random failure of some line in the code unless the user tries something really crazy (e.g., if they throw in a crystal structure to a composition object, a weird failure is fine).

Note that I am saying this because the featurize_dataframe() method includes an ignore_errors key, so I am assuming that if 999 of my entries are good and 1 entry is incompatible with a featurizer, I can still call featurize_dataframe and get data for the 999 and nan for the 1 without everything crashing. Subsequently, I can use pandas to fill nan values with 0 or some other value I like.

[ardunn]

For nan vs exception, I found the most efficient workflow for featurizing a lot (all ~60k MP mats) of (potentially) error prone data was to ignore_errors = True and filter mostly nan features with pandas nan threshold afterwards. But I can also see why throwing exceptions is good, in the case of accidentally putting in a structure for a comp or something, because if you had ignore_errors on, you would be essentially eliminating a good feature and not even know it.

Maybe in the future we can throw a warning to the user for each feature that has 90+% (or some parameter percentage) nan values. That way, even if you have a df with 300+ columns, you might not see which columns are mostly nan (and possibly due to a user error) unless you explicitly check, but you would get a warning saying "Look at feature X to make sure all these nan values are actually due to some crazy error, not due to accidentally putting in an invalid argument."

[computron]

@ardunn @WardLT Should this issue be considered closed?

Note that for 90% nan errors, I think we should just have a separate utility function for that OR just put that kind of filtering in our example notebooks using standard pandas tools (rather than build it into the featurize_dataframe).

[ardunn]

@computron This issue is closed. Also note my comments above are woefully outdated. I think the capabilities of matminer now (ie having both return_errors and ignore_errors) is sufficient for a user to figure out how a featurization went awry. I agree that the 90% thing should be taken care of by pandas once the df is returned.