The model (mostly) ignores every known sample except one
So there are extensions
$avg=\frac{1}{N} \sum_i knn_i(x)$
$wavg=\frac{1}{N} \sum_i \frac{knn_i(x)}{i}$