One-Hot Encoding

Encode a categorical column while preserving an ordered vocabulary. Categories in one fitting column are hashable and mutually sortable (for example, all strings or all integers).

Functions to implement

• get_categories(column): return the sorted unique categories.
• encode_value(value,categories): return one coordinate per supplied category. A known value has exactly one 1. An unknown value returns all zeros; do not change the vocabulary.
• one_hot_encode(column): fit categories with get_categories, encode each row, and return (encoded_matrix,categories). Empty input returns ([],[]).
• decode(one_hot_row,categories): return the category at the 1's index, or None for an all-zero row. Inputs have the correct width and at most one 1; other malformed rows are outside this exercise.

Fit and reuse

matrix, categories = one_hot_encode(["cat","dog","cat","bird"])
# categories = ["bird","cat","dog"]
# matrix = [[0,1,0],[0,0,1],[0,1,0],[1,0,0]]

decode([0,1,0],categories)  # "cat"

# Transform new values with the saved vocabulary:
[encode_value(v,categories) for v in ["dog","fox"]]
# [[0,0,1],[0,0,0]]
decode([0,0,0],categories)  # None

Do not call one_hot_encode again to transform a held-out batch: that fits new categories and can change both width and column meaning. Unknown encodings retain the saved width but cannot recover the original unseen value.