I am not sure what statistical test to use for comparing clustering on related data.

I have two sets of measurements on each person. I cluster each person by each measurement. I then measure how much the two coincide. I want to know how I would test for how significance the agreement of the clustering is.

For example I have 10 people and for each person I have two sets of attributes say vector A and Vector B both 10 of length 10.

I cluster the 10 people into 2 clusters using vector A and get something like

1 2 1 1 2 1 1 1 2 2

That is person one is assigned to cluster one, person two to cluster two person three to cluster one etc

I cluster the 10 examples into 2 clusters using vector B and get something like

1 2 1 1 2 1 1 1 1 1

These two clustering agree to 80% but how significant is this? What is the formula for calculating this? Is there a variant for more than 2 clusters? would it make sense if there were a different number of clusters for each measurement or if some samples only had one set of measurements?

Thank you for any thoughts :)

  • 1
    $\begingroup$ I am not sure to understand precisely what you are after or what “significant” means in this context but you might be interested in measures of “inter-rater agreement”. $\endgroup$ – Gala Jul 10 '13 at 9:37

Look up the literature on external evaluation of clusterings.

There are at least 30 measures to compute a similarity of two clusterings. Maybe one of these measures will give you significance values.


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.