KNN classifier labels

[manisci]

Hi,

Thank you for the great documentation and scripts. I think there is something missing in the method "_find_nearest" in the class "KNNClassifier". I will try to illustrate it with an example, let's assume the labels for the closest neighbors, sorted by closeness is nearest_label = [1 0 0 0 1] and we are using uniform distances so nearest_distance is [1 1 1 1 1]. We expect the prediction to be 0, since there are 3 zeros, but using the code in "_find_nearest", we would get 1.

import numpy as np
nearest_labels = np.array([1, 0, 0, 0, 1])
nearest_distances = np.array([1, 1, 1, 1, 1])
labels_distances = np.vstack((nearest_labels, nearest_distances));
labels_distances = labels_distances[:, labels_distances[0, :].argsort()]
i = np.nonzero(np.diff(labels_distances[0, :]))[0] + 1
i = np.insert(i, 0, 0)
label_scores = np.add.reduceat(labels_distances[1, :], I)
max_labels = nearest_labels[np.argmax(label_scores)]
print (max_labels)

In particular, I don't understand the reasoning for this line

max_labels = nearest_labels[np.argmax(label_scores)]

I believe this should be replaced by

ordered_labels = np.array(sorted(list(set(nearest_labels))))
max_labels = ordered_labels [np.argmax(label_scores)]

This would give the expected result of 0.

[eonu]

Hey @manisci thanks for pointing this out, glad to see another user.

Will take a look!

[eonu]

Fixed in #187.