At the end of the day, the targets we predict in the MNIST dataset are just numbers between 0 and 9. So, we can alternatively use a regressor algorithm for the same problem. In this case, our predictions will not be integers anymore, but rather floats. Training the regressor isn't much different from training the classifier:
from sklearn.neighbors import KNeighborsRegressor
clf = KNeighborsRegressor(n_neighbors=3, metric='euclidean')
clf.fit(x_train, y_train)
y_test_pred = clf.predict(x_test)
Here are some of the incorrectly made predictions:
The first item's three nearest neighbors are 3, 3, and 5. So, the regressor used their mean (3.67) as the prediction. The second and third items' neighbors are 8, 9, 8 and 7, 9, 7, respectively. Remember to round these predictions and convert them into integers if you want to use a classifier's evaluation metric to evaluate...
