Simplify BaseEncoder mapping
julia-kraus opened this issue · 1 comments
Current State
The class variable mapping
of the BaseEncoder class is a dictionary with the following structure
a dictionary with the following item:
- 'col' -> column name (str)
- 'mapping' -> pd.Series containing the mapping from category to encoding
Suggestion
Avoid pandas data frames and just use dictionaries instead. These are faster, easier to read and easier to manipulate. Mappings could be chained more easily. Moreover, many of the encoders, they are converted to dict anyways.
Use a nested dictionary of the structure {colname: {'cat_a': mapping_val_a, 'cat_b': mapping_val_b, 'cat_c': mapping_val_c ,...)
Actually the whole encoder could just yield the mapping dictionary, because the column names can be retrieved by mapping.keys().
This would really be helpful for better development and readability, but would be work to update and test all subclasses. I think it is unnecessary that users can supply both dictionaries and pd.Series as mappings because they can run pd.Series.to_dict()
themselves.
Hi Julia,
thanks for your suggestion. I think you're right and we should do it. This would be a major simplification but also a major change in the API. Maybe we can put it in the backlog for version 3.0 together with more simplifications and restructuring that we've been discussing lately