Running locally on a Jupyter notebook and using the MNIST dataset (28k entries, 28x28 pixels per image, the following takes 27 seconds.

from sklearn.neighbors import KNeighborsClassifierknn_clf = KNeighborsClassifier(n_jobs=1)
knn_clf.fit(pixels, labels)

However, the following takes 1722 seconds, in other words ~64 times longer:

from sklearn.model_selection import cross_val_predict
y_train_pred = cross_val_predict(knn_clf, pixels, labels, cv = 3, n_jobs=1)

My naive understanding is that cross_val_predict with cv=3 is doing 3-fold cross validation, so I'd expect it to fit the model 3 times, and so take at least ~3 times longer, but I don't see why it would take 64x!

To check if it was something specific to my environment, I ran the same in a Colab notebook - the difference was less extreme (15x), but still way above the ~3x I expected:

What am I missing? Why is cross_val_predict so much slower than just fitting the model?

In case it matters, I'm running scikit-learn 0.20.2.

KNN is also called as lazy algorithm because during fitting it does nothing but saves the input data, specifically there is no learning at all.

During predict is the actual distance calculation happens for each test datapoint. Hence, you could understand that when using cross_val_predict, KNN has to predict on the validation data points, which makes the computation time higher!