A Confusion Matrix in One Pass

Before any per-class metric there is a confusion matrix, and building one with two nested loops over classes is the answer that gets you asked to try again.

Implement confusion_matrix(y_true, y_pred, n_classes). Given two integer arrays of the same length whose values are class indices in [0, n_classes), return an (n_classes, n_classes) integer array where entry [i, j] is the number of samples whose true class is i and whose predicted class is j.

y_true = [0, 1, 1, 2]
y_pred = [0, 1, 2, 2]

         pred 0  pred 1  pred 2
true 0        1       0       0
true 1        0       1       1
true 2        0       0       1

Rows are true classes and columns are predictions. A class that appears in neither array still gets its own row and column of zeros: the matrix is always (n_classes, n_classes).

Do it in one pass with no Python loops. The trick is that a pair of small integers is itself a single integer.