[python] Stratified Train/Test-split in scikit-learn

I need to split my data into a training set (75%) and test set (25%). I currently do that with the code below:

X, Xt, userInfo, userInfo_train = sklearn.cross_validation.train_test_split(X, userInfo)   

However, I'd like to stratify my training dataset. How do I do that? I've been looking into the StratifiedKFold method, but doesn't let me specifiy the 75%/25% split and only stratify the training dataset.

In addition to the accepted answer by @Andreas Mueller, just want to add that as @tangy mentioned above:

StratifiedShuffleSplit most closely resembles train_test_split(stratify = y) with added features of:

1. stratify by default
2. by specifying n_splits, it repeatedly splits the data

You can simply do it with train_test_split() method available in Scikit learn:

from sklearn.model_selection import train_test_split 
train, test = train_test_split(X, test_size=0.25, stratify=X['YOUR_COLUMN_LABEL']) 

I have also prepared a short GitHub Gist which shows how stratify option works:

https://gist.github.com/SHi-ON/63839f3a3647051a180cb03af0f7d0d9

Updating @tangy answer from above to the current version of scikit-learn: 0.23.2 (StratifiedShuffleSplit documentation).

from sklearn.model_selection import StratifiedShuffleSplit

n_splits = 1  # We only want a single split in this case
sss = StratifiedShuffleSplit(n_splits=n_splits, test_size=0.25, random_state=0)

for train_index, test_index in sss.split(X, y):
    X_train, X_test = X[train_index], X[test_index]
    y_train, y_test = y[train_index], y[test_index]

As such, it is desirable to split the dataset into train and test sets in a way that preserves the same proportions of examples in each class as observed in the original dataset.

This is called a stratified train-test split.

We can achieve this by setting the “stratify” argument to the y component of the original dataset. This will be used by the train_test_split() function to ensure that both the train and test sets have the proportion of examples in each class that is present in the provided “y” array.

#train_size is 1 - tst_size - vld_size
tst_size=0.15
vld_size=0.15

X_train_test, X_valid, y_train_test, y_valid = train_test_split(df.drop(y, axis=1), df.y, test_size = vld_size, random_state=13903) 

X_train_test_V=pd.DataFrame(X_train_test)
X_valid=pd.DataFrame(X_valid)

X_train, X_test, y_train, y_test = train_test_split(X_train_test, y_train_test, test_size=tst_size, random_state=13903)

TL;DR : Use StratifiedShuffleSplit with test_size=0.25

Scikit-learn provides two modules for Stratified Splitting:

1. StratifiedKFold : This module is useful as a direct k-fold cross-validation operator: as in it will set up n_folds training/testing sets such that classes are equally balanced in both.

Heres some code(directly from above documentation)

>>> skf = cross_validation.StratifiedKFold(y, n_folds=2) #2-fold cross validation
>>> len(skf)
2
>>> for train_index, test_index in skf:
...    print("TRAIN:", train_index, "TEST:", test_index)
...    X_train, X_test = X[train_index], X[test_index]
...    y_train, y_test = y[train_index], y[test_index]
...    #fit and predict with X_train/test. Use accuracy metrics to check validation performance

2. StratifiedShuffleSplit : This module creates a single training/testing set having equally balanced(stratified) classes. Essentially this is what you want with the n_iter=1. You can mention the test-size here same as in train_test_split

Code:

>>> sss = StratifiedShuffleSplit(y, n_iter=1, test_size=0.5, random_state=0)
>>> len(sss)
1
>>> for train_index, test_index in sss:
...    print("TRAIN:", train_index, "TEST:", test_index)
...    X_train, X_test = X[train_index], X[test_index]
...    y_train, y_test = y[train_index], y[test_index]
>>> # fit and predict with your classifier using the above X/y train/test

Here's an example for continuous/regression data (until this issue on GitHub is resolved).

min = np.amin(y)
max = np.amax(y)

# 5 bins may be too few for larger datasets.
bins     = np.linspace(start=min, stop=max, num=5)
y_binned = np.digitize(y, bins, right=True)

X_train, X_test, y_train, y_test = train_test_split(
    X, 
    y, 
    stratify=y_binned
)

• Where start is min and stop is max of your continuous target.
• If you don't set right=True then it will more or less make your max value a separate bin and your split will always fail because too few samples will be in that extra bin.