StratifiedKFold与Kfold在shuffle后的主要区别

不shuffle时KFold按照各样本原来的顺序简单分折，而StratifiedKFold分层抽样

from sklearn.model_selection import KFold,StratifiedKFold
X = np.array([
    [1, 2, 3, 4],
    [11, 12, 13, 14],
    [21, 22, 23, 24],
    [31, 32, 33, 34],
    [41, 42, 43, 44],
    [51, 52, 53, 54],
    [61, 62, 63, 64],
    [71, 72, 73, 74],
    [71, 74, 73, 74],
])

y = np.array([1, 1, 0, 0, 1, 1, 0, 0, 0])
folder = KFold(n_splits=4, shuffle=False)
sfolder = StratifiedKFold(n_splits=4, shuffle=False)

for train, test in folder.split(X, y):
    print('Train: %s | test: %s' % (train, test))
    print(" ")
    print(pd.Series(y[train]).value_counts())
    print(pd.Series(y[test]).value_counts())
print('-------------------------------------')
for train, test in sfolder.split(X, y):
    print('Train: %s | test: %s' % (train, test))
    print(" ")
    print(pd.Series(y[train]).value_counts())
    print(pd.Series(y[test]).value_counts())

Train: [3 4 5 6 7 8] | test: [0 1 2]
 
0    4
1    2
dtype: int64
1    2
0    1
dtype: int64
Train: [0 1 2 5 6 7 8] | test: [3 4]
 
0    4
1    3
dtype: int64
1    1
0    1
dtype: int64
Train: [0 1 2 3 4 7 8] | test: [5 6]
 
0    4
1    3
dtype: int64
1    1
0    1
dtype: int64
Train: [0 1 2 3 4 5 6] | test: [7 8]
 
1    4
0    3
dtype: int64
0    2
dtype: int64
-------------------------------------
Train: [1 4 5 6 7 8] | test: [0 2 3]
 
1    3
0    3
dtype: int64
0    2
1    1
dtype: int64
Train: [0 2 3 4 5 7 8] | test: [1 6]
 
0    4
1    3
dtype: int64
1    1
0    1
dtype: int64
Train: [0 1 2 3 5 6 8] | test: [4 7]
 
0    4
1    3
dtype: int64
1    1
0    1
dtype: int64
Train: [0 1 2 3 4 6 7] | test: [5 8]
 
0    4
1    3
dtype: int64
1    1
0    1
dtype: int64

shuffle时，KFold随机打乱后分折，StratifiedKFold打乱但仍保持分层抽样的原则

folder = KFold(n_splits=4, shuffle=True,random_state=1)
sfolder = StratifiedKFold(n_splits=4, shuffle=True,random_state=1)

注：random_state用来保存随机的状态，以便实验复现。默认random_state=None可以看作0，即如果不设置random_state，下次的打乱情况还是一样的。

Train: [0 1 3 4 5 7] | test: [2 6 8]
 
1    4
0    2
dtype: int64
0    3
dtype: int64
Train: [0 2 3 4 5 6 8] | test: [1 7]
 
0    4
1    3
dtype: int64
1    1
0    1
dtype: int64
Train: [1 2 3 5 6 7 8] | test: [0 4]
 
0    5
1    2
dtype: int64
1    2
dtype: int64
Train: [0 1 2 4 6 7 8] | test: [3 5]
 
0    4
1    3
dtype: int64
1    1
0    1
dtype: int64
-------------------------------------
Train: [0 1 3 5 6 7] | test: [2 4 8]
 
1    3
0    3
dtype: int64
0    2
1    1
dtype: int64
Train: [0 1 2 4 6 7 8] | test: [3 5]
 
0    4
1    3
dtype: int64
1    1
0    1
dtype: int64
Train: [0 2 3 4 5 6 8] | test: [1 7]
 
0    4
1    3
dtype: int64
1    1
0    1
dtype: int64
Train: [1 2 3 4 5 7 8] | test: [0 6]
 
0    4
1    3
dtype: int64
1    1
0    1
dtype: int64