category

Categorical transformation for data science

Installation

pip install works for this library.

pip install category

Single Category

>>> from category import Category
>>> book = Category(['a', 'b', 'c', 'Category_d', 'e', 'f', 'g', 'h', 'i', 'j'], pad_mst = False)
>>> book.i2c[2]
'c'

>>> book.c2i[['Category_d','f']]
array([3, 5])

You can set pad_mst to True to handle the missing token

>>> from category import Category
>>> book = Category(['a', 'b', 'c', 'Category_d', 'e', 'f', 'g', 'h', 'i', 'j'], pad_mst = True)
>>> book.i2c[2] # the 1st token is the missing token, not 'a' any more
'b'
>>> book.c2i[['Stranger','Category_d','Unknown','f']]
array([0, 4, 0, 6])

Multi-Category

>>> from category import (Category, MultiCategory)
>>> cates = list(f"category{i}" for i in range(1000))
>>> multi_cate = MultiCategory(Category(cates, pad_mst = True))
>>> multi_cate.string_to_index("category42, category108")
array([42, 108])

You can also try to convert a list of strings, containing multicategorical info (which the data input is frequently used in tabular data), to nhot encoded array, and back

>>> nhot = multi_cate.batch_strings_to_nhot(["category42, category108","category999"])
>>> multi_cate.nhot_to_list(nhot)[0]
["category42", "category108"]

Performance

The running speed of this library mostly depends on python dictionary and numpy operations. Though python is a 'slow' language, such application is pretty fast, and not easy to improve using other language.

Here we compare the this library with the Rust implementation

References

• GitHub
• PyPI package
• Rust implementation
• Used in Tai-Chi engine, a verstile user-friendly deep learning library