0
$\begingroup$

I want to do clustering based on the similarity kernel $K(x,y)$, but on know from my ground truth labels that for 10-20% of the data the most similar data samples are not in the same class (based on the used metric).

So does it mean that clustering for this data using a that similarity kernel would be a wrong choice? By the way, i think that'd be a typical problem for all real datasets as not always the closest data samples are in the same class, but there are many similarity based clustering methods tried and tested and also published!

$\endgroup$
1
$\begingroup$

Some examples having their closest neighbor be of a different class is almost unavoidable if the clusters are close together. However this can be an indication that k-nearest neighbors is not the algorithm to use. Have you tried k-means to see how well that works?

$\endgroup$

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.