[FEA] Deprecate label_encoding method in favor of categorical columns

[lastephey]

Dear cuDF developers,

Is your feature request related to a problem? Please describe.

My feature request is for cuDF to add the ability for the label_encoding function to encode string data. My specific use-case is to encode string values as ints.

Describe the solution you'd like

I would like to use the label_encoding function to encode some values that are currently strings as ints. My larger goal is to use these data in CuPy's histogram2d, but since CuPy does not yet support string arrays, I was hoping that label encoding would help me get around this problem.

Describe alternatives you've considered

I have also tried the cuDF replace function for this, but it looks like it doesn't support encoding strings as ints either:

to_replace and value should be of same types,got to_replace dtype: object and value dtype: int64

I'd be equally happy with replace supporting the ability to replace objects of different types.

Additional context

If there are better ways to accomplish this task, I would be grateful for your suggestions.

Thank you very much,
Laurie

[davidwendt]

What does the string data look like? You may be able to use the subword_tokenize function to assign integers to words.

If you want to replace an entire string row with an integer you could use replace with strings that contain integers and then cast that result to int64

>>> import cudf
>>> s = cudf.Series(['a','b','a','c'])
>>> s.str.replace(['a','b','c'],['1','2','3'], regex=False).astype('int64')
0    1
1    2
2    1
3    3
dtype: int64

[beckernick]

In addition to @davidwendt , you could do this in a couple of other ways to avoid enumerating your categories (the former probably being preferred method in pandas):

import cudf
​
s = cudf.Series(["a","b","c"])
s.astype("category").cat.codes
0    0
1    1
2    2
dtype: uint8

import cudf
from cuml.preprocessing import LabelEncoder
​
s = cudf.Series(["a","b","c"])
​
le = LabelEncoder()
le.fit(s)
le.transform(s)
0    0
1    1
2    2
dtype: uint8

cc @shwina perhaps we should consider deprecating label_encode as the core functionality is generally covered by categorical columns and cuML's LabelEncoder. From the git blame it's from the early pyGDF days years ago. I don't think this method is supported in pandas.

[lastephey]

Thank you @davidwendt and @beckernick for your suggestions. I think they all sound like they will do what I need. The strings I need to replace are all values in a single column.

I haven't had time to try these out yet but I just wanted to say thank you and that I really appreciate your quick answers and advice.

[lastephey]

Ok, I've had some time to read about and test the options you suggested.

1. subword_tokenize would definitely work for what I need to do, but there's some overhead to set this up.
2. Using replace as @davidwendt suggested works well, but it does result in this warning for me using cuDF 21.08.00a:

/opt/miniconda3/lib/python3.8/site-packages/cudf/core/column/string.py:945: UserWarning: `n` parameter is not supported when `pat` and `repl` are list-like inputs
  warnings.warn(

3. Using categorial data is the simplest and my favorite. I didn't even know about this functionality within Pandas, so thank you very much for the pointer.
4. the cuML solution also seems fine, although in my case I'm not using cuML so I don't want to add this extra dependency unless I have to.

Thank you both again for your help and suggestions. I really appreciate it. My question is resolved. Should I close this, or would you like me to keep it open for issue tracking?

[beckernick]

Glad this is resolved. Let's keep this open and edit the title to be a discussion around deprecating label_encoding.

[vyasr]

This was done in #9535 and #9289