VotingClassifier: Different Feature Sets

Question

I have two different feature sets (so, with same number of rows and the labels are the same), in my case DataFrames:

df1:

| A | B | C |
-------------
| 1 | 4 | 2 |
| 1 | 4 | 8 |
| 2 | 1 | 1 |
| 2 | 3 | 0 |
| 3 | 2 | 5 |

df2:

| E | F |
---------
| 6 | 1 |
| 1 | 3 |
| 8 | 1 |
| 2 | 8 |
| 5 | 2 |

labels:

| labels |
----------
|    5   |
|    5   |
|    1   |
|    7   |
|    3   |

I want to use them to train a VotingClassifier. But the fitting step only allows to specify a single feature set. Goal is to fit clf1 with df1 and clf2 with df2.

eclf = VotingClassifier(estimators=[('df1-clf', clf1), ('df2-clf', clf2)], voting='soft')
eclf.fit(...)

How should I proceed with this kind of situation? Is there any easy solution?

Answer 1

Its pretty easy to make custom functions to do what you want to achieve.

Import the prerequisites:

import numpy as np
from sklearn.preprocessing import LabelEncoder

def fit_multiple_estimators(classifiers, X_list, y, sample_weights = None):

    # Convert the labels `y` using LabelEncoder, because the predict method is using index-based pointers
    # which will be converted back to original data later.
    le_ = LabelEncoder()
    le_.fit(y)
    transformed_y = le_.transform(y)

    # Fit all estimators with their respective feature arrays
    estimators_ = [clf.fit(X, y) if sample_weights is None else clf.fit(X, y, sample_weights) for clf, X in zip([clf for _, clf in classifiers], X_list)]

    return estimators_, le_


def predict_from_multiple_estimator(estimators, label_encoder, X_list, weights = None):

    # Predict 'soft' voting with probabilities

    pred1 = np.asarray([clf.predict_proba(X) for clf, X in zip(estimators, X_list)])
    pred2 = np.average(pred1, axis=0, weights=weights)
    pred = np.argmax(pred2, axis=1)

    # Convert integer predictions to original labels:
    return label_encoder.inverse_transform(pred)

The logic is taken from VotingClassifier source.

Now test the above methods. First get some data:

from sklearn.datasets import load_iris
data = load_iris()
X = data.data
y = []

#Convert int classes to string labels
for x in data.target:
    if x==0:
        y.append('setosa')
    elif x==1:
        y.append('versicolor')
    else:
        y.append('virginica')

Split the data into train and test:

from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y)

Divide the X into different feature datas:

X_train1, X_train2 = X_train[:,:2], X_train[:,2:]
X_test1, X_test2 = X_test[:,:2], X_test[:,2:]

X_train_list = [X_train1, X_train2]
X_test_list = [X_test1, X_test2]

Get list of classifiers:

from sklearn.neighbors import KNeighborsClassifier
from sklearn.svm import SVC

# Make sure the number of estimators here are equal to number of different feature datas
classifiers = [('knn',  KNeighborsClassifier(3)),
    ('svc', SVC(kernel="linear", C=0.025, probability=True))]

Fit the classifiers with the data:

fitted_estimators, label_encoder = fit_multiple_estimators(classifiers, X_train_list, y_train)

Predict using the test data:

y_pred = predict_from_multiple_estimator(fitted_estimators, label_encoder, X_test_list)

Get accuracy of predictions:

from sklearn.metrics import accuracy_score
print(accuracy_score(y_test, y_pred))

Feel free to ask if any doubt.

Answer 2

To use as much as sklearn tools as possible, I find following way more appealing.

from sklearn.base import TransformerMixin, BaseEstimator
import numpy as np
from sklearn.pipeline import Pipeline
from sklearn.linear_model import LogisticRegression
from sklearn.svm import SVC
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from sklearn.ensemble import VotingClassifier

######################
# custom transformer for sklearn pipeline
class ColumnExtractor(TransformerMixin, BaseEstimator):
    def __init__(self, cols):
        self.cols = cols

    def transform(self, X):
        col_list = []
        for c in self.cols:
            col_list.append(X[:, c:c+1])
        return np.concatenate(col_list, axis=1)

    def fit(self, X, y=None):
        return self

######################
# processing data
data = load_iris()
X = data.data
y = data.target
X_train, X_test, y_train, y_test = train_test_split(X, y)

######################
# fit clf1 with df1
pipe1 = Pipeline([
    ('col_extract', ColumnExtractor( cols=range(0,2) )), # selecting features 0 and 1 (df1) to be used with LR (clf1)
    ('clf', LogisticRegression())
    ])

pipe1.fit(X_train, y_train) # sanity check
pipe1.score(X_test,y_test) # sanity check
# output: 0.6842105263157895

######################
# fit clf2 with df2
pipe2 = Pipeline([
    ('col_extract', ColumnExtractor( cols=range(2,4) )), # selecting features 2 and 3 (df2) to be used with SVC (clf2)
    ('clf', SVC(probability=True))
    ])

pipe2.fit(X_train, y_train) # sanity check
pipe2.score(X_test,y_test) # sanity check
# output: 0.9736842105263158

######################
# ensemble/voting classifier where clf1 fitted with df1 and clf2 fitted with df2
eclf = VotingClassifier(estimators=[('df1-clf1', pipe1), ('df2-clf2', pipe2)], voting='soft', weights= [1, 0.5])
eclf.fit(X_train, y_train)
eclf.score(X_test,y_test)
# output: 0.9473684210526315