I get the same pattern for 3 different indices: Silhouettes, Dunn and Connectivity- as the number of clusters increases, the score decreases. I am using several clustering methods and several distance metrics, this pattern is the same for each combination. I can't figure out why.


Silhouettes scores for increasing number of clusters

  • 2
    $\begingroup$ Your data may simply not exhibit any obvious clusters, i.e., clustering may not make sense at all. $\endgroup$ – Stephan Kolassa Apr 24 '15 at 13:18

First of all, if you didn't preprocess your data well, if it doesn't have a lot of clusters, or if there are outliers, the clustering algorithms may fail to work well. So maybe that's happening for you.

Also, measures such as sum-of-squares by definition should decrease with increasing number of clusters. The optimum is often achieved when every object is it's own cluster - a useless result. Thus, do not rely on the measures too much. They can guide you to interesing results (when the measure changed rapidly) and they can detect similar results (little change, not interesting) but in the end, inspect the result manually, don't rely on an evaluation number to tell you what is good - it cannot.


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.