# Sequential Feature Selector

Implementation of sequential feature algorithms (SFAs) -- greedy search algorithms -- that have been developed as a suboptimal solution to the computationally often not feasible exhaustive search.

from mlxtend.feature_selection import SequentialFeatureSelector

## Overview

Sequential feature selection algorithms are a family of greedy search algorithms that are used to reduce an initial d-dimensional feature space to a k-dimensional feature subspace where k < d. The motivation behind feature selection algorithms is to automatically select a subset of features that is most relevant to the problem. The goal of feature selection is two-fold: We want to improve the computational efficiency and reduce the generalization error of the model by removing irrelevant features or noise. A wrapper approach such as sequential feature selection is especially useful if embedded feature selection -- for example, a regularization penalty like LASSO -- is not applicable.

In a nutshell, SFAs remove or add one feature at the time based on the classifier performance until a feature subset of the desired size k is reached. There are 4 different flavors of SFAs available via the SequentialFeatureSelector:

1. Sequential Forward Selection (SFS)
2. Sequential Backward Selection (SBS)
3. Sequential Forward Floating Selection (SFFS)
4. Sequential Backward Floating Selection (SBFS)

The floating variants, SFFS and SBFS, can be considered as extensions to the simpler SFS and SBS algorithms. The floating algorithms have an additional exclusion or inclusion step to remove features once they were included (or excluded), so that a larger number of feature subset combinations can be sampled. It is important to emphasize that this step is conditional and only occurs if the resulting feature subset is assessed as "better" by the criterion function after removal (or addition) of a particular feature. Furthermore, I added an optional check to skip the conditional exclusion steps if the algorithm gets stuck in cycles.

How is this different from Recursive Feature Elimination (RFE) -- e.g., as implemented in sklearn.feature_selection.RFE? RFE is computationally less complex using the feature weight coefficients (e.g., linear models) or feature importance (tree-based algorithms) to eliminate features recursively, whereas SFSs eliminate (or add) features based on a user-defined classifier/regression performance metric.

The SFAs are outlined in pseudo code below:

### Sequential Forward Selection (SFS)

Input: $Y = \{y_1, y_2, ..., y_d\}$

• The SFS algorithm takes the whole $d$-dimensional feature set as input.

Output: $X_k = \{x_j \; | \;j = 1, 2, ..., k; \; x_j \in Y\}$, where $k = (0, 1, 2, ..., d)$

• SFS returns a subset of features; the number of selected features $k$, where $k < d$, has to be specified a priori.

Initialization: $X_0 = \emptyset$, $k = 0$

• We initialize the algorithm with an empty set $\emptyset$ ("null set") so that $k = 0$ (where $k$ is the size of the subset).

Step 1 (Inclusion):

$x^+ = \text{ arg max } J(x_k + x), \text{ where } x \in Y - X_k$
$X_{k+1} = X_k + x^+$
$k = k + 1$
Go to Step 1

• in this step, we add an additional feature, $x^+$, to our feature subset $X_k$.
• $x^+$ is the feature that maximizes our criterion function, that is, the feature that is associated with the best classifier performance if it is added to $X_k$.
• We repeat this procedure until the termination criterion is satisfied.

Termination: $k = p$

• We add features from the feature subset $X_k$ until the feature subset of size $k$ contains the number of desired features $p$ that we specified a priori.

### Sequential Backward Selection (SBS)

Input: the set of all features, $Y = \{y_1, y_2, ..., y_d\}$

• The SBS algorithm takes the whole feature set as input.

Output: $X_k = \{x_j \; | \;j = 1, 2, ..., k; \; x_j \in Y\}$, where $k = (0, 1, 2, ..., d)$

• SBS returns a subset of features; the number of selected features $k$, where $k < d$, has to be specified a priori.

Initialization: $X_0 = Y$, $k = d$

• We initialize the algorithm with the given feature set so that the $k = d$.

Step 1 (Exclusion):

$x^- = \text{ arg max } J(x_k - x), \text{ where } x \in X_k$
$X_{k-1} = X_k - x^-$
$k = k - 1$
Go to Step 1

• In this step, we remove a feature, $x^-$ from our feature subset $X_k$.
• $x^-$ is the feature that maximizes our criterion function upon re,oval, that is, the feature that is associated with the best classifier performance if it is removed from $X_k$.
• We repeat this procedure until the termination criterion is satisfied.

Termination: $k = p$

• We add features from the feature subset $X_k$ until the feature subset of size $k$ contains the number of desired features $p$ that we specified a priori.

### Sequential Backward Floating Selection (SBFS)

Input: the set of all features, $Y = \{y_1, y_2, ..., y_d\}$

• The SBFS algorithm takes the whole feature set as input.

Output: $X_k = \{x_j \; | \;j = 1, 2, ..., k; \; x_j \in Y\}$, where $k = (0, 1, 2, ..., d)$

• SBFS returns a subset of features; the number of selected features $k$, where $k < d$, has to be specified a priori.

Initialization: $X_0 = Y$, $k = d$

• We initialize the algorithm with the given feature set so that the $k = d$.

Step 1 (Exclusion):

$x^- = \text{ arg max } J(x_k - x), \text{ where } x \in X_k$
$X_{k-1} = X_k - x^-$
$k = k - 1$
Go to Step 2

• In this step, we remove a feature, $x^-$ from our feature subset $X_k$.
• $x^-$ is the feature that maximizes our criterion function upon re,oval, that is, the feature that is associated with the best classifier performance if it is removed from $X_k$.

Step 2 (Conditional Inclusion):

$x^+ = \text{ arg max } J(x_k + x), \text{ where } x \in Y - X_k$
if J(x_k + x) > J(x_k + x):
$X_{k+1} = X_k + x^+$
$k = k + 1$
Go to Step 1

• In Step 2, we search for features that improve the classifier performance if they are added back to the feature subset. If such features exist, we add the feature $x^+$ for which the performance improvement is maximized. If $k = 2$ or an improvement cannot be made (i.e., such feature $x^+$ cannot be found), go back to step 1; else, repeat this step.

Termination: $k = p$

• We add features from the feature subset $X_k$ until the feature subset of size $k$ contains the number of desired features $p$ that we specified a priori.

### Sequential Forward Floating Selection (SFFS)

Input: the set of all features, $Y = \{y_1, y_2, ..., y_d\}$

• The SFFS algorithm takes the whole feature set as input, if our feature space consists of, e.g. 10, if our feature space consists of 10 dimensions (d = 10).

Output: a subset of features, $X_k = \{x_j \; | \;j = 1, 2, ..., k; \; x_j \in Y\}$, where $k = (0, 1, 2, ..., d)$

• The returned output of the algorithm is a subset of the feature space of a specified size. E.g., a subset of 5 features from a 10-dimensional feature space (k = 5, d = 10).

Initialization: $X_0 = Y$, $k = d$

• We initialize the algorithm with an empty set ("null set") so that the k = 0 (where k is the size of the subset)

Step 1 (Inclusion):

$x^+ = \text{ arg max } J(x_k + x), \text{ where } x \in Y - X_k$
$X_{k+1} = X_k + x^+$
$k = k + 1$
Go to Step 2

Step 2 (Conditional Exclusion):

$x^- = \text{ arg max } J(x_k - x), \text{ where } x \in X_k$
$if \; J(x_k - x) > J(x_k - x)$:
$X_{k-1} = X_k - x^-$
$k = k - 1$
Go to Step 1

• In step 1, we include the feature from the feature space that leads to the best performance increase for our feature subset (assessed by the criterion function). Then, we go over to step 2
• In step 2, we only remove a feature if the resulting subset would gain an increase in performance. If $k = 2$ or an improvement cannot be made (i.e., such feature $x^+$ cannot be found), go back to step 1; else, repeat this step.

• Steps 1 and 2 are repeated until the Termination criterion is reached.

Termination: stop when k equals the number of desired features

## Example 1 - A simple Sequential Forward Selection example

Initializing a simple classifier from scikit-learn:

from sklearn.neighbors import KNeighborsClassifier

X = iris.data
y = iris.target
knn = KNeighborsClassifier(n_neighbors=4)


We start by selection the "best" 3 features from the Iris dataset via Sequential Forward Selection (SFS). Here, we set forward=True and floating=False. By choosing cv=0, we don't perform any cross-validation, therefore, the performance (here: 'accuracy') is computed entirely on the training set.

from mlxtend.feature_selection import SequentialFeatureSelector as SFS

sfs1 = SFS(knn,
k_features=3,
forward=True,
floating=False,
verbose=2,
scoring='accuracy',
cv=0)

sfs1 = sfs1.fit(X, y)

[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   4 out of   4 | elapsed:    0.0s finished

[2018-05-06 12:49:16] Features: 1/3 -- score: 0.96[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   3 out of   3 | elapsed:    0.0s finished

[2018-05-06 12:49:16] Features: 2/3 -- score: 0.973333333333[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   2 out of   2 | elapsed:    0.0s finished

[2018-05-06 12:49:16] Features: 3/3 -- score: 0.973333333333


Via the subsets_ attribute, we can take a look at the selected feature indices at each step:

sfs1.subsets_

{1: {'avg_score': 0.95999999999999996,
'cv_scores': array([ 0.96]),
'feature_idx': (3,),
'feature_names': ('3',)},
2: {'avg_score': 0.97333333333333338,
'cv_scores': array([ 0.97333333]),
'feature_idx': (2, 3),
'feature_names': ('2', '3')},
3: {'avg_score': 0.97333333333333338,
'cv_scores': array([ 0.97333333]),
'feature_idx': (1, 2, 3),
'feature_names': ('1', '2', '3')}}


Note that the 'feature_names' entry is simply a string representation of the 'feature_idx' in this case. Optionally, we can provide custom feature names via the fit method's custom_feature_names parameter:

feature_names = ('sepal length', 'sepal width', 'petal length', 'petal width')
sfs1 = sfs1.fit(X, y, custom_feature_names=feature_names)
sfs1.subsets_

[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   4 out of   4 | elapsed:    0.0s finished

[2018-05-06 12:49:16] Features: 1/3 -- score: 0.96[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   3 out of   3 | elapsed:    0.0s finished

[2018-05-06 12:49:16] Features: 2/3 -- score: 0.973333333333[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   2 out of   2 | elapsed:    0.0s finished

[2018-05-06 12:49:16] Features: 3/3 -- score: 0.973333333333

{1: {'avg_score': 0.95999999999999996,
'cv_scores': array([ 0.96]),
'feature_idx': (3,),
'feature_names': ('petal width',)},
2: {'avg_score': 0.97333333333333338,
'cv_scores': array([ 0.97333333]),
'feature_idx': (2, 3),
'feature_names': ('petal length', 'petal width')},
3: {'avg_score': 0.97333333333333338,
'cv_scores': array([ 0.97333333]),
'feature_idx': (1, 2, 3),
'feature_names': ('sepal width', 'petal length', 'petal width')}}


Furthermore, we can access the indices of the 3 best features directly via the k_feature_idx_ attribute:

sfs1.k_feature_idx_

(1, 2, 3)


And similarly, to obtain the names of these features, given that we provided an argument to the custom_feature_names parameter, we can refer to the sfs1.k_feature_names_ attribute:

sfs1.k_feature_names_

('sepal width', 'petal length', 'petal width')


Finally, the prediction score for these 3 features can be accesses via k_score_:

sfs1.k_score_

0.97333333333333338


## Example 2 - Toggling between SFS, SBS, SFFS, and SBFS

Using the forward and floating parameters, we can toggle between SFS, SBS, SFFS, and SBFS as shown below. Note that we are performing (stratified) 4-fold cross-validation for more robust estimates in contrast to Example 1. Via n_jobs=-1, we choose to run the cross-validation on all our available CPU cores.

# Sequential Forward Selection
sfs = SFS(knn,
k_features=3,
forward=True,
floating=False,
scoring='accuracy',
cv=4,
n_jobs=-1)
sfs = sfs.fit(X, y)

print('\nSequential Forward Selection (k=3):')
print(sfs.k_feature_idx_)
print('CV Score:')
print(sfs.k_score_)

###################################################

# Sequential Backward Selection
sbs = SFS(knn,
k_features=3,
forward=False,
floating=False,
scoring='accuracy',
cv=4,
n_jobs=-1)
sbs = sbs.fit(X, y)

print('\nSequential Backward Selection (k=3):')
print(sbs.k_feature_idx_)
print('CV Score:')
print(sbs.k_score_)

###################################################

# Sequential Forward Floating Selection
sffs = SFS(knn,
k_features=3,
forward=True,
floating=True,
scoring='accuracy',
cv=4,
n_jobs=-1)
sffs = sffs.fit(X, y)

print('\nSequential Forward Floating Selection (k=3):')
print(sffs.k_feature_idx_)
print('CV Score:')
print(sffs.k_score_)

###################################################

# Sequential Backward Floating Selection
sbfs = SFS(knn,
k_features=3,
forward=False,
floating=True,
scoring='accuracy',
cv=4,
n_jobs=-1)
sbfs = sbfs.fit(X, y)

print('\nSequential Backward Floating Selection (k=3):')
print(sbfs.k_feature_idx_)
print('CV Score:')
print(sbfs.k_score_)

Sequential Forward Selection (k=3):
(1, 2, 3)
CV Score:
0.972756410256

Sequential Backward Selection (k=3):
(1, 2, 3)
CV Score:
0.972756410256

Sequential Forward Floating Selection (k=3):
(1, 2, 3)
CV Score:
0.972756410256

Sequential Backward Floating Selection (k=3):
(1, 2, 3)
CV Score:
0.972756410256


In this simple scenario, selecting the best 3 features out of the 4 available features in the Iris set, we end up with similar results regardless of which sequential selection algorithms we used.

## Example 3 - Visualizing the results in DataFrames

For our convenience, we can visualize the output from the feature selection in a pandas DataFrame format using the get_metric_dict method of the SequentialFeatureSelector object. The columns std_dev and std_err represent the standard deviation and standard errors of the cross-validation scores, respectively.

Below, we see the DataFrame of the Sequential Forward Selector from Example 2:

import pandas as pd
pd.DataFrame.from_dict(sfs.get_metric_dict()).T

avg_score ci_bound cv_scores feature_idx feature_names std_dev std_err
1 0.952991 0.0660624 [0.974358974359, 0.948717948718, 0.88888888888... (3,) (3,) 0.0412122 0.0237939
2 0.959936 0.0494801 [0.974358974359, 0.948717948718, 0.91666666666... (2, 3) (2, 3) 0.0308676 0.0178214
3 0.972756 0.0315204 [0.974358974359, 1.0, 0.944444444444, 0.972222... (1, 2, 3) (1, 2, 3) 0.0196636 0.0113528

Now, let's compare it to the Sequential Backward Selector:

pd.DataFrame.from_dict(sbs.get_metric_dict()).T

avg_score ci_bound cv_scores feature_idx feature_names std_dev std_err
3 0.972756 0.0315204 [0.974358974359, 1.0, 0.944444444444, 0.972222... (1, 2, 3) (1, 2, 3) 0.0196636 0.0113528
4 0.952991 0.0372857 [0.974358974359, 0.948717948718, 0.91666666666... (0, 1, 2, 3) (0, 1, 2, 3) 0.0232602 0.0134293

We can see that both SFS and SBFS found the same "best" 3 features, however, the intermediate steps where obviously different.

The ci_bound column in the DataFrames above represents the confidence interval around the computed cross-validation scores. By default, a confidence interval of 95% is used, but we can use different confidence bounds via the confidence_interval parameter. E.g., the confidence bounds for a 90% confidence interval can be obtained as follows:

pd.DataFrame.from_dict(sbs.get_metric_dict(confidence_interval=0.90)).T

avg_score ci_bound cv_scores feature_idx feature_names std_dev std_err
3 0.972756 0.0242024 [0.974358974359, 1.0, 0.944444444444, 0.972222... (1, 2, 3) (1, 2, 3) 0.0196636 0.0113528
4 0.952991 0.0286292 [0.974358974359, 0.948717948718, 0.91666666666... (0, 1, 2, 3) (0, 1, 2, 3) 0.0232602 0.0134293

## Example 4 - Plotting the results

After importing the little helper function plotting.plot_sequential_feature_selection, we can also visualize the results using matplotlib figures.

from mlxtend.plotting import plot_sequential_feature_selection as plot_sfs
import matplotlib.pyplot as plt

sfs = SFS(knn,
k_features=4,
forward=True,
floating=False,
scoring='accuracy',
verbose=2,
cv=5)

sfs = sfs.fit(X, y)

fig1 = plot_sfs(sfs.get_metric_dict(), kind='std_dev')

plt.ylim([0.8, 1])
plt.title('Sequential Forward Selection (w. StdDev)')
plt.grid()
plt.show()

[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   4 out of   4 | elapsed:    0.0s finished

[2018-05-06 12:49:18] Features: 1/4 -- score: 0.96[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   3 out of   3 | elapsed:    0.0s finished

[2018-05-06 12:49:18] Features: 2/4 -- score: 0.966666666667[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   2 out of   2 | elapsed:    0.0s finished

[2018-05-06 12:49:18] Features: 3/4 -- score: 0.953333333333[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s finished

[2018-05-06 12:49:18] Features: 4/4 -- score: 0.973333333333


## Example 6 - Using Pandas DataFrames

import pandas as pd
from sklearn.neighbors import KNeighborsClassifier
from mlxtend.feature_selection import SequentialFeatureSelector as SFS

X = iris.data
y = iris.target
knn = KNeighborsClassifier(n_neighbors=4)

sfs1 = SFS(knn,
k_features=3,
forward=True,
floating=False,
scoring='accuracy',
cv=0)

X_df = pd.DataFrame(X, columns=['sepal len', 'petal len',
'sepal width', 'petal width'])

sepal len petal len sepal width petal width
0 5.1 3.5 1.4 0.2
1 4.9 3.0 1.4 0.2
2 4.7 3.2 1.3 0.2
3 4.6 3.1 1.5 0.2
4 5.0 3.6 1.4 0.2

Also, the target array, y, can be optionally be cast as a Series:

y_series = pd.Series(y)

0    0
1    0
2    0
3    0
4    0
dtype: int64

sfs1 = sfs1.fit(X_df, y_series)


Note that the only difference of passing a pandas DataFrame as input is that the sfs1.subsets_ array will now contain a new column,

sfs1.subsets_

{1: {'avg_score': 0.95999999999999996,
'cv_scores': array([ 0.96]),
'feature_idx': (3,),
'feature_names': ('petal width',)},
2: {'avg_score': 0.97333333333333338,
'cv_scores': array([ 0.97333333]),
'feature_idx': (2, 3),
'feature_names': ('sepal width', 'petal width')},
3: {'avg_score': 0.97333333333333338,
'cv_scores': array([ 0.97333333]),
'feature_idx': (1, 2, 3),
'feature_names': ('petal len', 'sepal width', 'petal width')}}


In mlxtend version >= 0.13 pandas DataFrames are supported as feature inputs to the SequentianFeatureSelector instead of NumPy arrays or other NumPy-like array types.

## Example 5 - Sequential Feature Selection for Regression

Similar to the classification examples above, the SequentialFeatureSelector also supports scikit-learn's estimators for regression.

from sklearn.linear_model import LinearRegression

X, y = boston.data, boston.target

lr = LinearRegression()

sfs = SFS(lr,
k_features=13,
forward=True,
floating=False,
scoring='neg_mean_squared_error',
cv=10)

sfs = sfs.fit(X, y)
fig = plot_sfs(sfs.get_metric_dict(), kind='std_err')

plt.title('Sequential Forward Selection (w. StdErr)')
plt.grid()
plt.show()


## Example 6 -- Using the Selected Feature Subset For Making New Predictions

# Initialize the dataset

from sklearn.neighbors import KNeighborsClassifier
from sklearn.model_selection import train_test_split

X, y = iris.data, iris.target
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.33, random_state=1)

knn = KNeighborsClassifier(n_neighbors=4)

# Select the "best" three features via
# 5-fold cross-validation on the training set.

from mlxtend.feature_selection import SequentialFeatureSelector as SFS

sfs1 = SFS(knn,
k_features=3,
forward=True,
floating=False,
scoring='accuracy',
cv=5)
sfs1 = sfs1.fit(X_train, y_train)

print('Selected features:', sfs1.k_feature_idx_)

Selected features: (1, 2, 3)

# Generate the new subsets based on the selected features
# Note that the transform call is equivalent to
# X_train[:, sfs1.k_feature_idx_]

X_train_sfs = sfs1.transform(X_train)
X_test_sfs = sfs1.transform(X_test)

# Fit the estimator using the new feature subset
# and make a prediction on the test data
knn.fit(X_train_sfs, y_train)
y_pred = knn.predict(X_test_sfs)

# Compute the accuracy of the prediction
acc = float((y_test == y_pred).sum()) / y_pred.shape[0]
print('Test set accuracy: %.2f %%' % (acc * 100))

Test set accuracy: 96.00 %


## Example 7 -- Sequential Feature Selection and GridSearch

# Initialize the dataset

from sklearn.neighbors import KNeighborsClassifier
from sklearn.model_selection import train_test_split

X, y = iris.data, iris.target
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.33, random_state=1)


Use scikit-learn's GridSearch to tune the hyperparameters inside and outside the SequentialFeatureSelector:

from sklearn.model_selection import GridSearchCV
from sklearn.pipeline import Pipeline
from mlxtend.feature_selection import SequentialFeatureSelector as SFS
import mlxtend

knn = KNeighborsClassifier(n_neighbors=2)

sfs1 = SFS(estimator=knn,
k_features=3,
forward=True,
floating=False,
scoring='accuracy',
cv=5)

pipe = Pipeline([('sfs', sfs1),
('knn', knn)])

param_grid = [
{'sfs__k_features': [1, 2, 3, 4],
'sfs__estimator__n_neighbors': [1, 2, 3, 4]}
]

gs = GridSearchCV(estimator=pipe,
param_grid=param_grid,
scoring='accuracy',
n_jobs=1,
cv=5,
refit=False)

# run gridearch
gs = gs.fit(X_train, y_train)


... and the "best" parameters determined by GridSearch are ...

print("Best parameters via GridSearch", gs.best_params_)

Best parameters via GridSearch {'sfs__estimator__n_neighbors': 1, 'sfs__k_features': 3}


#### Obtaining the best k feature indices after GridSearch

If we are interested in the best k feature indices via SequentialFeatureSelection.k_feature_idx_, we have to initialize a GridSearchCV object with refit=True. Now, the grid search object will take the complete training dataset and the best parameters, which it found via cross-validation, to train the estimator pipeline.

gs = GridSearchCV(estimator=pipe,
param_grid=param_grid,
scoring='accuracy',
n_jobs=1,
cv=5,
refit=True)
gs = gs.fit(X_train, y_train)


After running the grid search, we can access the individual pipeline objects of the best_estimator_ via the steps attribute.

gs.best_estimator_.steps

[('sfs', SequentialFeatureSelector(clone_estimator=True, cv=5,
estimator=KNeighborsClassifier(algorithm='auto', leaf_size=30, metric='minkowski',
metric_params=None, n_jobs=1, n_neighbors=1, p=2,
weights='uniform'),
floating=False, forward=True, k_features=3, n_jobs=1,
pre_dispatch='2*n_jobs', scoring='accuracy', verbose=0)),
('knn',
KNeighborsClassifier(algorithm='auto', leaf_size=30, metric='minkowski',
metric_params=None, n_jobs=1, n_neighbors=2, p=2,
weights='uniform'))]


Via sub-indexing, we can then obtain the best-selected feature subset:

print('Best features:', gs.best_estimator_.steps[0][1].k_feature_idx_)

Best features: (0, 1, 3)


During cross-validation, this feature combination had a CV accuracy of:

print('Best score:', gs.best_score_)

Best score: 0.94

gs.best_params_

{'sfs__estimator__n_neighbors': 1, 'sfs__k_features': 3}


Alternatively, if we can set the "best grid search parameters" in our pipeline manually if we ran GridSearchCV with refit=False. It should yield the same results:

pipe.set_params(**gs.best_params_).fit(X_train, y_train)
print('Best features:', pipe.steps[0][1].k_feature_idx_)

Best features: (0, 1, 3)


## Example 8 -- Selecting the "best" feature combination in a k-range

If k_features is set to to a tuple (min_k, max_k) (new in 0.4.2), the SFS will now select the best feature combination that it discovered by iterating from k=1 to max_k (forward), or max_k to min_k (backward). The size of the returned feature subset is then within max_k to min_k, depending on which combination scored best during cross validation.

X.shape

(150, 4)

from mlxtend.feature_selection import SequentialFeatureSelector as SFS
from sklearn.neighbors import KNeighborsClassifier
from mlxtend.data import wine_data
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import make_pipeline

X, y = wine_data()
X_train, X_test, y_train, y_test= train_test_split(X, y,
stratify=y,
test_size=0.3,
random_state=1)

knn = KNeighborsClassifier(n_neighbors=2)

sfs1 = SFS(estimator=knn,
k_features=(3, 10),
forward=True,
floating=False,
scoring='accuracy',
cv=5)

pipe = make_pipeline(StandardScaler(), sfs1)

pipe.fit(X_train, y_train)

print('best combination (ACC: %.3f): %s\n' % (sfs1.k_score_, sfs1.k_feature_idx_))
print('all subsets:\n', sfs1.subsets_)
plot_sfs(sfs1.get_metric_dict(), kind='std_err');

best combination (ACC: 0.992): (0, 1, 2, 3, 6, 8, 9, 10, 11, 12)

all subsets:
{1: {'feature_idx': (6,), 'cv_scores': array([ 0.84615385,  0.6       ,  0.88      ,  0.79166667,  0.875     ]), 'avg_score': 0.7985641025641026, 'feature_names': ('6',)}, 2: {'feature_idx': (6, 9), 'cv_scores': array([ 0.92307692,  0.88      ,  1.        ,  0.95833333,  0.91666667]), 'avg_score': 0.93561538461538463, 'feature_names': ('6', '9')}, 3: {'feature_idx': (6, 9, 12), 'cv_scores': array([ 0.92307692,  0.92      ,  0.96      ,  1.        ,  0.95833333]), 'avg_score': 0.95228205128205123, 'feature_names': ('6', '9', '12')}, 4: {'feature_idx': (3, 6, 9, 12), 'cv_scores': array([ 0.96153846,  0.96      ,  0.96      ,  1.        ,  0.95833333]), 'avg_score': 0.96797435897435891, 'feature_names': ('3', '6', '9', '12')}, 5: {'feature_idx': (3, 6, 9, 10, 12), 'cv_scores': array([ 0.92307692,  0.96      ,  1.        ,  1.        ,  1.        ]), 'avg_score': 0.97661538461538466, 'feature_names': ('3', '6', '9', '10', '12')}, 6: {'feature_idx': (2, 3, 6, 9, 10, 12), 'cv_scores': array([ 0.92307692,  0.96      ,  1.        ,  0.95833333,  1.        ]), 'avg_score': 0.96828205128205125, 'feature_names': ('2', '3', '6', '9', '10', '12')}, 7: {'feature_idx': (0, 2, 3, 6, 9, 10, 12), 'cv_scores': array([ 0.92307692,  0.92      ,  1.        ,  1.        ,  1.        ]), 'avg_score': 0.96861538461538466, 'feature_names': ('0', '2', '3', '6', '9', '10', '12')}, 8: {'feature_idx': (0, 2, 3, 6, 8, 9, 10, 12), 'cv_scores': array([ 1.  ,  0.92,  1.  ,  1.  ,  1.  ]), 'avg_score': 0.98399999999999999, 'feature_names': ('0', '2', '3', '6', '8', '9', '10', '12')}, 9: {'feature_idx': (0, 2, 3, 6, 8, 9, 10, 11, 12), 'cv_scores': array([ 1.  ,  0.92,  1.  ,  1.  ,  1.  ]), 'avg_score': 0.98399999999999999, 'feature_names': ('0', '2', '3', '6', '8', '9', '10', '11', '12')}, 10: {'feature_idx': (0, 1, 2, 3, 6, 8, 9, 10, 11, 12), 'cv_scores': array([ 1.  ,  0.96,  1.  ,  1.  ,  1.  ]), 'avg_score': 0.99199999999999999, 'feature_names': ('0', '1', '2', '3', '6', '8', '9', '10', '11', '12')}}


## Example 9 -- Using other cross-validation schemes

In addition to standard k-fold and stratified k-fold, other cross validation schemes can be used with SequentialFeatureSelector. For example, GroupKFold or LeaveOneOut cross-validation from scikit-learn.

#### Using GroupKFold with SequentialFeatureSelector

from mlxtend.feature_selection import SequentialFeatureSelector as SFS
from sklearn.neighbors import KNeighborsClassifier
from mlxtend.data import iris_data
from sklearn.model_selection import GroupKFold
import numpy as np

X, y = iris_data()
groups = np.arange(len(y)) // 10
print('groups: {}'.format(groups))

groups: [ 0  0  0  0  0  0  0  0  0  0  1  1  1  1  1  1  1  1  1  1  2  2  2  2  2
2  2  2  2  2  3  3  3  3  3  3  3  3  3  3  4  4  4  4  4  4  4  4  4  4
5  5  5  5  5  5  5  5  5  5  6  6  6  6  6  6  6  6  6  6  7  7  7  7  7
7  7  7  7  7  8  8  8  8  8  8  8  8  8  8  9  9  9  9  9  9  9  9  9  9
10 10 10 10 10 10 10 10 10 10 11 11 11 11 11 11 11 11 11 11 12 12 12 12 12
12 12 12 12 12 13 13 13 13 13 13 13 13 13 13 14 14 14 14 14 14 14 14 14 14]


Calling the split() method of a scikit-learn cross-validator object will return a generator that yields train, test splits.

cv_gen = GroupKFold(4).split(X, y, groups)
cv_gen

<generator object _BaseKFold.split at 0x1a1a041200>


The cv parameter of SequentialFeatureSelector must be either an int or an iterable yielding train, test splits. This iterable can be constructed by passing the train, test split generator to the built-in list() function.

cv = list(cv_gen)

knn = KNeighborsClassifier(n_neighbors=2)
sfs = SFS(estimator=knn,
k_features=2,
scoring='accuracy',
cv=cv)

sfs.fit(X, y)

print('best combination (ACC: %.3f): %s\n' % (sfs.k_score_, sfs.k_feature_idx_))

best combination (ACC: 0.940): (2, 3)


## Example 10 - Working with pandas DataFrames

Optionally, we can also use pandas DataFrames and pandas Series as input to the fit function. In this case, the column names of the pandas DataFrame will be used as feature names. However, note that if custom_feature_names are provided in the fit function, these custom_feature_names take precedence over the DataFrame column-based feature names.

import pandas as pd
from sklearn.neighbors import KNeighborsClassifier

col_names = ('sepal length', 'sepal width',
'petal length', 'petal width')
X_df = pd.DataFrame(iris.data, columns=col_names)
y_series = pd.Series(iris.target)
knn = KNeighborsClassifier(n_neighbors=4)

from mlxtend.feature_selection import SequentialFeatureSelector as SFS

sfs1 = SFS(knn,
k_features=3,
forward=True,
floating=False,
verbose=2,
scoring='accuracy',
cv=0)

sfs1 = sfs1.fit(X_df, y_series)

[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   4 out of   4 | elapsed:    0.0s finished

[2018-05-06 12:49:29] Features: 1/3 -- score: 0.96[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   3 out of   3 | elapsed:    0.0s finished

[2018-05-06 12:49:29] Features: 2/3 -- score: 0.973333333333[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.0s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   2 out of   2 | elapsed:    0.0s finished

[2018-05-06 12:49:29] Features: 3/3 -- score: 0.973333333333

sfs1.k_feature_names_

('sepal width', 'petal length', 'petal width')


# API

SequentialFeatureSelector(estimator, k_features=1, forward=True, floating=False, verbose=0, scoring=None, cv=5, n_jobs=1, pre_dispatch='2n_jobs', clone_estimator=True)*

Sequential Feature Selection for Classification and Regression.

Parameters

• estimator : scikit-learn classifier or regressor

• k_features : int or tuple or str (default: 1)

Number of features to select, where k_features < the full feature set. New in 0.4.2: A tuple containing a min and max value can be provided, and the SFS will consider return any feature combination between min and max that scored highest in cross-validtion. For example, the tuple (1, 4) will return any combination from 1 up to 4 features instead of a fixed number of features k. New in 0.8.0: A string argument "best" or "parsimonious". If "best" is provided, the feature selector will return the feature subset with the best cross-validation performance. If "parsimonious" is provided as an argument, the smallest feature subset that is within one standard error of the cross-validation performance will be selected.

• forward : bool (default: True)

Forward selection if True, backward selection otherwise

• floating : bool (default: False)

Adds a conditional exclusion/inclusion if True.

• verbose : int (default: 0), level of verbosity to use in logging.

If 0, no output, if 1 number of features in current set, if 2 detailed logging i ncluding timestamp and cv scores at step.

• scoring : str, callable, or None (default: None)

If None (default), uses 'accuracy' for sklearn classifiers and 'r2' for sklearn regressors. If str, uses a sklearn scoring metric string identifier, for example {accuracy, f1, precision, recall, roc_auc} for classifiers, {'mean_absolute_error', 'mean_squared_error'/'neg_mean_squared_error', 'median_absolute_error', 'r2'} for regressors. If a callable object or function is provided, it has to be conform with sklearn's signature scorer(estimator, X, y); see http://scikit-learn.org/stable/modules/generated/sklearn.metrics.make_scorer.html for more information.

• cv : int (default: 5)

Integer or iterable yielding train, test splits. If cv is an integer and estimator is a classifier (or y consists of integer class labels) stratified k-fold. Otherwise regular k-fold cross-validation is performed. No cross-validation if cv is None, False, or 0.

• n_jobs : int (default: 1)

The number of CPUs to use for evaluating different feature subsets in parallel. -1 means 'all CPUs'.

• pre_dispatch : int, or string (default: '2*n_jobs')

Controls the number of jobs that get dispatched during parallel execution if n_jobs > 1 or n_jobs=-1. Reducing this number can be useful to avoid an explosion of memory consumption when more jobs get dispatched than CPUs can process. This parameter can be: None, in which case all the jobs are immediately created and spawned. Use this for lightweight and fast-running jobs, to avoid delays due to on-demand spawning of the jobs An int, giving the exact number of total jobs that are spawned A string, giving an expression as a function of n_jobs, as in 2*n_jobs

• clone_estimator : bool (default: True)

Clones estimator if True; works with the original estimator instance if False. Set to False if the estimator doesn't implement scikit-learn's set_params and get_params methods. In addition, it is required to set cv=0, and n_jobs=1.

Attributes

• k_feature_idx_ : array-like, shape = [n_predictions]

Feature Indices of the selected feature subsets.

• k_feature_names_ : array-like, shape = [n_predictions]

Feature names of the selected feature subsets. If pandas DataFrames are used in the fit method, the feature names correspond to the column names. Otherwise, the feature names are string representation of the feature array indices. New in v 0.13.0.

• k_score_ : float

Cross validation average score of the selected subset.

• subsets_ : dict

A dictionary of selected feature subsets during the sequential selection, where the dictionary keys are the lengths k of these feature subsets. The dictionary values are dictionaries themselves with the following keys: 'feature_idx' (tuple of indices of the feature subset) 'feature_names' (tuple of feature names of the feat. subset) 'cv_scores' (list individual cross-validation scores) 'avg_score' (average cross-validation score) Note that if pandas DataFrames are used in the fit method, the 'feature_names' correspond to the column names. Otherwise, the feature names are string representation of the feature array indices. The 'feature_names' is new in v 0.13.0.

Examples

For usage examples, please see http://rasbt.github.io/mlxtend/user_guide/feature_selection/SequentialFeatureSelector/

### Methods

fit(X, y, custom_feature_names=None, fit_params)

Perform feature selection and learn model from training data.

Parameters

• X : {array-like, sparse matrix}, shape = [n_samples, n_features]

Training vectors, where n_samples is the number of samples and n_features is the number of features. New in v 0.13.0: pandas DataFrames are now also accepted as argument for X.

• y : array-like, shape = [n_samples]

Target values. New in v 0.13.0: pandas DataFrames are now also accepted as argument for y.

• custom_feature_names : None or tuple (default: tuple)

Custom feature names for self.k_feature_names and self.subsets_[i]['feature_names']. (new in v 0.13.0)

• fit_params : dict of string -> object, optional

Parameters to pass to to the fit method of classifier.

Returns

• self : object

fit_transform(X, y, fit_params)

Fit to training data then reduce X to its most important features.

Parameters

• X : {array-like, sparse matrix}, shape = [n_samples, n_features]

Training vectors, where n_samples is the number of samples and n_features is the number of features. New in v 0.13.0: pandas DataFrames are now also accepted as argument for X.

• y : array-like, shape = [n_samples]

Target values. New in v 0.13.0: a pandas Series are now also accepted as argument for y.

• fit_params : dict of string -> object, optional

Parameters to pass to to the fit method of classifier.

Returns

Reduced feature subset of X, shape={n_samples, k_features}

get_metric_dict(confidence_interval=0.95)

Return metric dictionary

Parameters

• confidence_interval : float (default: 0.95)

A positive float between 0.0 and 1.0 to compute the confidence interval bounds of the CV score averages.

Returns

Dictionary with items where each dictionary value is a list with the number of iterations (number of feature subsets) as its length. The dictionary keys corresponding to these lists are as follows: 'feature_idx': tuple of the indices of the feature subset 'cv_scores': list with individual CV scores 'avg_score': of CV average scores 'std_dev': standard deviation of the CV score average 'std_err': standard error of the CV score average 'ci_bound': confidence interval bound of the CV score average

get_params(deep=True)

Get parameters for this estimator.

Parameters

• deep : boolean, optional

If True, will return the parameters for this estimator and contained subobjects that are estimators.

Returns

• params : mapping of string to any

Parameter names mapped to their values.

set_params(params)

Set the parameters of this estimator.

The method works on simple estimators as well as on nested objects (such as pipelines). The latter have parameters of the form <component>__<parameter> so that it's possible to update each component of a nested object.

Returns

self

transform(X)

Reduce X to its most important features.

Parameters

• X : {array-like, sparse matrix}, shape = [n_samples, n_features]

Training vectors, where n_samples is the number of samples and n_features is the number of features. New in v 0.13.0: pandas DataFrames are now also accepted as argument for X.

Returns

Reduced feature subset of X, shape={n_samples, k_features}