How to reverse sklearn.OneHotEncoder transform to recover original data?

Question

I encoded my categorical data using sklearn.OneHotEncoder and fed them to a random forest classifier. Everything seems to work and I got my predicted output bac

Answer 1

Since version 0.20 of scikit-learn, the active_features_ attribute of the OneHotEncoder class has been deprecated, so I suggest to rely on the categories_ attribute instead.

The below function can help you recover the original data from a matrix that has been one-hot encoded:

def reverse_one_hot(X, y, encoder):
    reversed_data = [{} for _ in range(len(y))]
    all_categories = list(itertools.chain(*encoder.categories_))
    category_names = ['category_{}'.format(i+1) for i in range(len(encoder.categories_))]
    category_lengths = [len(encoder.categories_[i]) for i in range(len(encoder.categories_))]

    for row_index, feature_index in zip(*X.nonzero()):
        category_value = all_categories[feature_index]
        category_name = get_category_name(feature_index, category_names, category_lengths)
        reversed_data[row_index][category_name] = category_value
        reversed_data[row_index]['target'] = y[row_index]

    return reversed_data


def get_category_name(index, names, lengths):

    counter = 0
    for i in range(len(lengths)):
        counter += lengths[i]
        if index < counter:
            return names[i]
    raise ValueError('The index is higher than the number of categorical values')

To test it, I have created a small data set that includes the ratings that users have given to users

data = [
    {'user_id': 'John', 'item_id': 'The Matrix', 'rating': 5},
    {'user_id': 'John', 'item_id': 'Titanic', 'rating': 1},
    {'user_id': 'John', 'item_id': 'Forrest Gump', 'rating': 2},
    {'user_id': 'John', 'item_id': 'Wall-E', 'rating': 2},
    {'user_id': 'Lucy', 'item_id': 'The Matrix', 'rating': 5},
    {'user_id': 'Lucy', 'item_id': 'Titanic', 'rating': 1},
    {'user_id': 'Lucy', 'item_id': 'Die Hard', 'rating': 5},
    {'user_id': 'Lucy', 'item_id': 'Forrest Gump', 'rating': 2},
    {'user_id': 'Lucy', 'item_id': 'Wall-E', 'rating': 2},
    {'user_id': 'Eric', 'item_id': 'The Matrix', 'rating': 2},
    {'user_id': 'Eric', 'item_id': 'Die Hard', 'rating': 3},
    {'user_id': 'Eric', 'item_id': 'Forrest Gump', 'rating': 5},
    {'user_id': 'Eric', 'item_id': 'Wall-E', 'rating': 4},
    {'user_id': 'Diane', 'item_id': 'The Matrix', 'rating': 4},
    {'user_id': 'Diane', 'item_id': 'Titanic', 'rating': 3},
    {'user_id': 'Diane', 'item_id': 'Die Hard', 'rating': 5},
    {'user_id': 'Diane', 'item_id': 'Forrest Gump', 'rating': 3},
]

data_frame = pandas.DataFrame(data)
data_frame = data_frame[['user_id', 'item_id', 'rating']]
ratings = data_frame['rating']
data_frame.drop(columns=['rating'], inplace=True)

If we are building a prediction model, we have to remember to delete the dependent variable (in this case the rating) from the DataFrame before we encode it.

ratings = data_frame['rating']
data_frame.drop(columns=['rating'], inplace=True)

Then we proceed to do the encoding

ohc = OneHotEncoder()
encoded_data = ohc.fit_transform(data_frame)
print(encoded_data)

Which results in:

  (0, 2)    1.0
  (0, 6)    1.0
  (1, 2)    1.0
  (1, 7)    1.0
  (2, 2)    1.0
  (2, 5)    1.0
  (3, 2)    1.0
  (3, 8)    1.0
  (4, 3)    1.0
  (4, 6)    1.0
  (5, 3)    1.0
  (5, 7)    1.0
  (6, 3)    1.0
  (6, 4)    1.0
  (7, 3)    1.0
  (7, 5)    1.0
  (8, 3)    1.0
  (8, 8)    1.0
  (9, 1)    1.0
  (9, 6)    1.0
  (10, 1)   1.0
  (10, 4)   1.0
  (11, 1)   1.0
  (11, 5)   1.0
  (12, 1)   1.0
  (12, 8)   1.0
  (13, 0)   1.0
  (13, 6)   1.0
  (14, 0)   1.0
  (14, 7)   1.0
  (15, 0)   1.0
  (15, 4)   1.0
  (16, 0)   1.0
  (16, 5)   1.0

After encoding the we can reverse using the reverse_one_hot function we defined above, like this:

reverse_data = matrix_utils.reverse_one_hot(encoded_data, ratings, ohc)
print(pandas.DataFrame(reverse_data))

Which gives us:

   category_1    category_2  target
0        John    The Matrix       5
1        John       Titanic       1
2        John  Forrest Gump       2
3        John        Wall-E       2
4        Lucy    The Matrix       5
5        Lucy       Titanic       1
6        Lucy      Die Hard       5
7        Lucy  Forrest Gump       2
8        Lucy        Wall-E       2
9        Eric    The Matrix       2
10       Eric      Die Hard       3
11       Eric  Forrest Gump       5
12       Eric        Wall-E       4
13      Diane    The Matrix       4
14      Diane       Titanic       3
15      Diane      Die Hard       5
16      Diane  Forrest Gump       3

Answer 2

How to one-hot encode

See https://stackoverflow.com/a/42874726/562769

import numpy as np
nb_classes = 6
data = [[2, 3, 4, 0]]

def indices_to_one_hot(data, nb_classes):
    """Convert an iterable of indices to one-hot encoded labels."""
    targets = np.array(data).reshape(-1)
    return np.eye(nb_classes)[targets]

How to reverse

def one_hot_to_indices(data):
    indices = []
    for el in data:
        indices.append(list(el).index(1))
    return indices


hot = indices_to_one_hot(orig_data, nb_classes)
indices = one_hot_to_indices(hot)

print(orig_data)
print(indices)

gives:

[[2, 3, 4, 0]]
[2, 3, 4, 0]