For nearest-neighbors regression, it is plausible to increase the number of neighbors used to predict $f(x)$ when there are many data points near $x$.
What is a good algorithm for varying the number of neighbors used?
$\begingroup$
$\endgroup$
1
-
2$\begingroup$ Here's a paper describing a kNN algorithm with dynamic $k$: link.springer.com/chapter/10.1007/978-3-540-75185-4_7 $\endgroup$– zachdjSep 9, 2019 at 16:00
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
"A good algorithm for varying the number of neighbors" is probably quite data-dependent. But there is, for example, the fixed radius approach.
answered Oct 9, 2019 at 22:30
Ben Reiniger - on strike♦Ben Reiniger - on strike
10.9k2 gold badges14 silver badges52 bronze badges
$\begingroup$
$\endgroup$
1
The most common method that's used is "The Elbow Method." You decide your neighbours based on the error rate. You can use GridSearch to determine the ideal number of neighbours to be used. This gives you an overview on how to do it.
-
2$\begingroup$ As I tried to convey in my question, I don't want to used a fixed k for all data points but to vary k based on the density of data points. $\endgroup$ Sep 9, 2019 at 14:01