Why is cross_val_score different to when I calculate it manually?

Question

Here is the reproducible example code:

from numpy import mean
from sklearn.datasets import make_classification
from sklearn.model_selection import cross_validate
from sklearn.model_selection import StratifiedShuffleSplit
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import balanced_accuracy_score

# define dataset
X, y = make_classification(n_samples=1000, weights = [0.3,0.7], n_features=100, n_informative=75, random_state=0)
# define the model
model = RandomForestClassifier(n_estimators=10, random_state=0)
# evaluate the model
n_splits=10
cv = StratifiedShuffleSplit(n_splits, random_state=0)
n_scores = cross_validate(model, X, y, scoring='balanced_accuracy', cv=cv, n_jobs=-1, error_score='raise')
# report performance
print('Accuracy: %0.4f' % (mean(n_scores['test_score'])))

bal_acc_sum = []
for train_index, test_index in cv.split(X,y):
    model.fit(X[train_index], y[train_index])                                      
    bal_acc_sum.append(balanced_accuracy_score(model.predict(X[test_index]),y[test_index]))

print(f"Accuracy: %0.4f" % (mean(bal_acc_sum)))

Result:

Accuracy: 0.6737
Accuracy: 0.7113

The results for my self calculated accuracy is always higher than the one cross-validation gives me. But it should be the same or am I missing something? Same metric, same split (KFold brings same result), same fixed model (other models behave identically), same random state, but different results?

Answer 1

It is because, in your manual calculation, you have flipped the order of arguments in balanced_accuracy_score , which matters - it should be (y_true, y_pred) ( docs ).

Changing this, your manual calculation becomes:

bal_acc_sum = []
for train_index, test_index in cv.split(X,y):
    model.fit(X[train_index], y[train_index])                                      
    bal_acc_sum.append(balanced_accuracy_score(y[test_index], model.predict(X[test_index])))  # change order of arguments here

print(f"Accuracy: %0.4f" % (mean(bal_acc_sum)))

Result:

Accuracy: 0.6737

And

import numpy as np
np.all(bal_acc_sum==n_scores['test_score'])
# True