How to one hot encode variant length features?

Given a list of variant length features:

features = [
    ['f1', 'f2', 'f3'],
    ['f2', 'f4', 'f5', 'f6'],
    ['f1', 'f2']
]

where each sample has variant number of features and the feature dtype is str and already one hot.

In order to use feature selection utilities of sklearn, I have to convert the features to a 2D-array which looks like:

    f1  f2  f3  f4  f5  f6
s1   1   1   1   0   0   0
s2   0   1   0   1   1   1
s3   1   1   0   0   0   0

How could I achieve it via sklearn or numpy?

Solution 1:

You can use MultiLabelBinarizer present in scikit which is specifically used for doing this.

Code for your example:

features = [
            ['f1', 'f2', 'f3'],
            ['f2', 'f4', 'f5', 'f6'],
            ['f1', 'f2']
           ]
from sklearn.preprocessing import MultiLabelBinarizer
mlb = MultiLabelBinarizer()
new_features = mlb.fit_transform(features)

Output:

array([[1, 1, 1, 0, 0, 0],
       [0, 1, 0, 1, 1, 1],
       [1, 1, 0, 0, 0, 0]])

This can also be used in a pipeline, along with other feature_selection utilities.

Solution 2:

Here's one approach with NumPy methods and outputting as pandas dataframe -

import numpy as np
import pandas as pd

lens = list(map(len, features))
N = len(lens)
unq, col = np.unique(np.concatenate(features),return_inverse=1)
row = np.repeat(np.arange(N), lens)
out = np.zeros((N,len(unq)),dtype=int)
out[row,col] = 1

indx = ['s'+str(i+1) for i in range(N)]
df_out = pd.DataFrame(out, columns=unq, index=indx)

Sample input, output -

In [80]: features
Out[80]: [['f1', 'f2', 'f3'], ['f2', 'f4', 'f5', 'f6'], ['f1', 'f2']]

In [81]: df_out
Out[81]: 
    f1  f2  f3  f4  f5  f6
s1   1   1   1   0   0   0
s2   0   1   0   1   1   1
s3   1   1   0   0   0   0