4
$\begingroup$

The k-means++ algorithm provides a technique to choose the initial k seeds for the k-means algorithm. It does this by sampling the next point according to a multinomial distribution over the unchosen points (where the probability of a point being chosen as the next center is proportional to $D(x)^2$ with $D(x)$ being the distance of the point $x$ to its nearest center).

The point with the largest distance has the greatest probability of being chosen, but why can I not choose this point every time? What advantage do I gain by being 'fuzzy' with my seed selection?

$\endgroup$
4
$\begingroup$

You gain in theoretical guarantees on the solutions: the solution found by k-means initialized this way is close to the correct k-means solution (in expectation) with a known constant, cf. these slides for example.

With the method you mention (which was previously used in the literature), you can build some configurations where it behaves badly (think of a point on a separating hyperplane but far far away) for sure (since deterministic).

$\endgroup$
0
$\begingroup$
  1. K-means can get stuck in local minima.

Because of this, it is a best practise to run multiple times and keep the best result (best by SSQ).

If you always choose the farthest point, you will get the same result every time. So you do want some randomness!

  1. The farthest point is not the best centroid candidate. It usually is too far out.

The point has the highest probability of being chosen, but the average point chosen is much closer. If there are 10 points at distance^2 10, and one point at distance^2 11, then the algorithm is more likely to choose one of the cluster points than that last outlier.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.