GridSearchCV fit call: ValueError: n_splits=5 cannot be greater than the number of members in each class

Question

I'm trying to call GridSearchCV fit and get error. But can't understand this error.

Input:

X_train.info()

<class 'pandas.core.frame.DataFrame'>
Int64Index: 6000 entries, 9761 to 7270
Data columns (total 4 columns):
 #   Column  Non-Null Count  Dtype
---  ------  --------------  -----
 0   a       6000 non-null   int64
 1   b       6000 non-null   int64
 2   c       6000 non-null   int64
 3   d       6000 non-null   int64
dtypes: int64(4)
memory usage: 234.4 KB

y_train.info()

<class 'pandas.core.series.Series'>
Int64Index: 6000 entries, 9761 to 7270
Series name: result
Non-Null Count  Dtype
--------------  -----
6000 non-null   int64
dtypes: int64(1)
memory usage: 93.8 KB

GridSearchCV call:

parameters = {
    "max_depth": [1, 2, 3],
}

cv = GridSearchCV(
    DecisionTreeClassifier(), 
    parameters, 
    cv=5,
    verbose=1,
)

cv.fit(X_train, y_train)

And I receive error:

ValueError: n_splits=5 cannot be greater than the number of members in each class.

Why? Please explain this error. Thanks.

Answer 1

With cv=5 you ask for a five-fold stratified split to be performed. This means the percentage of each class is preserved for each split. The smallest class in your data appears to have less than 5 members so a stratified split is not possible: You need at least 5 members in each class so that at least one can be assigned to every split. You can take a look at the class distribution with y_train.value_counts().

A quick solution would be to reduce the number of folds, e.g.

cv = GridSearchCV(
    DecisionTreeClassifier(), 
    parameters, 
    cv=2,
    verbose=1,
)

Alternatively you could perform a non-stratified five-fold split like so

from sklearn.model_selection import KFold

kf5 = KFold(n_splits=5)
cv = GridSearchCV(
    DecisionTreeClassifier(), 
    parameters, 
    cv=kf5,
    verbose=1,
)

For more details on stratified splitting and other forms of cross validation see the sklearn user guide.