API/BUG: awkward syntax to add categories to a Categorical

[jreback]

from SO

• the .add_categories should be able to take an Index/ndarray, ATM it must be converted to a list
• should be a keyword for something like: add any additional categories that I am passing to you (even if the current ones are duplicates, just add these to the end). e.g.

In [147]: s = pd.Categorical(list('aabbcd'))

In [148]: s2 = list('aabbcdefg')

In [149]: s
Out[149]: 
[a, a, b, b, c, d]
Categories (4, object): [a, b, c, d]

In [150]: s.add_categories(s.categories.sym_diff(Index(s2)).tolist())
Out[150]: 
[a, a, b, b, c, d]
Categories (7, object): [a, b, c, d, e, f, g]

I would ideally just like to say:

s.add_categories(s2, take_new=True)

(maybe not the best keyword, but something like this)

[jreback]

cc @JanSchulz

[jankatins]

I'm against s.add_categories(s2, take_new=True): how should the whole things be ordered? If it is an ordered cat, the order has meaning and a user will run into surprises, no matter how good we document that :-(

A bad example for "append the difference": take two series, one ["a","b"] and one ["a","c"] and then use the above with ["a","b","c"] -> you get two different categories, one ["a","b","c"] and one ["a","c","b"].

We do have set_categories(new), which is more well defined (i.e. take the order from new, even if they are currently unused).

I think accepting ndarray/ndex would be good, but with the same restraints as the current list.

[jreback]

cc @JanSchulz I am not actually suggesting anything new, here, just appending, where the order is not defined, except that it occurs AFTER the existing categories (as .add_categories does now), see #9929

I think this is a very natural thing to do, and pretty awkward to force the user to make sure that they are not adding duplicate categories. Of course they can use .set_categories, but then what is the point of .add_categories then?

[jankatins]

regarding "what's the point of add_cats": for the usecases, which categorical should IMO optimized for ("lickert scales" or "american states"), this method is mostly useless: I can't really think about a usecase where set_categories isn't better or at least "good enough" (but then, I only only used questionairs). As far as I remember, it was added in the discussion to round the API off (in addition to add_categories).

IMO this issue/SO question is another "workaround"/indicator for the "categoricals as memory efficient strings": if you have a need for setting arbitrary length categories (e.g. s.unique()) without bothering about the order or what you actually add as categoricals then a memory efficient string type would be better. This is also nicely visible in the SO question: "I'm trying to reduce the size of ~300 csv files (about a billion rows) by replacing lengthy fields with shorter, categorical, values[...]"

[jreback]

see #9929

seems reasonable to me. @shoyer @jorisvandenbossche

@JanSchulz I not sure where you think this doesn't apply to Categoricals. I want to add a bunch of categories. Not saying they are in any particular order, just that we don't rewrite the existing category mappings. I think that is a reasonable API guarantee.

This is actually what we do in reality now, this just codifies it and makes it a convient to do.

[jankatins]

I just think that a report starting with "I'm trying to reduce the size..." is not something which should be used to influence the API design of Categorical but should be used to model a new MemoryEfficientString (which maybe has a common superclass with Categorical...).

Regarding the the usecases: I think that add_categories and remove_categories are not really needed if you only work with survey data (which is what I did and I've no experience with any other categorical data apart from the "I want more memory efficient strings" cases -> if there are more and they need the above methods, then my point is moot here), where you have something like 7 item lickert scales or 200 names of states: for them it makes as much (or more) sense to simple use set_categories as you start with the complete list of states or even want to have a defined order (lickert).

If you have a cat with ordered==True, then s.add_categories(s2, take_new=True) is IMO plain madness, as you can't be sure what the final order of the categories is (the a-b-c example above).

Even if you want "append in the end", a simple s.cat.add_categories(set(all_cats)-set(s.cat.categories)) is IMO not so terrible problematic to warrant a new kwarg (and I still think that s.cat.set_categories(all_cats) is better/easier/less error prone and IMO the more common case if you need a special order, where you anyway need to have the re-code/factorize).