Ordinal & Target Encoding

The problem, from first principles

01-one-hot-encoding turns a category into one column per value, which works until a column has thousands of values (a zip code, a product ID) and the table explodes. Two other encodings keep a single column. Ordinal encoding replaces each category by its rank when the categories have a natural order (small, medium, large). Target encoding replaces each category by the average of the label over the rows that have it, which packs the category's predictive signal into one number. Target encoding is also one of the easiest ways to leak the label into a feature by accident, so this question builds the safe versions.

From theory to code

Implement ordinal_encode(values, order) for ordered categories, then target_encode(categories, y, smoothing), which learns a smoothed mean per category, then apply_target_encoding(categories, mapping, default), which uses a learned mapping, then out_of_fold_target_encode(categories, y, n_folds, smoothing), which encodes every row using only rows from other folds. The signatures and docstrings are already in the editor.

Constraints

• values and categories are 1D arrays of hashable labels. y is a 1D float array of the same length. order is a list of labels from lowest to highest.
• ordinal_encode returns an integer array where each value is the position of its label in order. A label that is not in order becomes -1.
• target_encode returns a dict mapping each category to (n * category_mean + smoothing * global_mean) / (n + smoothing), where n is how many rows have that category and global_mean is the mean of all of y. With smoothing = 0 this is the plain category mean.
• apply_target_encoding returns a float array with one value per entry of categories. A category missing from mapping gets default.
• out_of_fold_target_encode splits the row indices 0 .. n-1 into n_folds contiguous, nearly equal blocks with np.array_split, in order, with no shuffling. For each block it learns the mapping from all other rows with target_encode and applies it to the block, using the global mean of the other rows as default. It returns a float array of length n.

Hints