The silhouette score in my case gives quite misleading results, any alternatives?

My data is a result of embedding words, which belong to one of the 20+ classes. I want to measure the "clusteredness" of words belonging to the same label. Often times words of the same label cluster into a 2-3 well defined blobs, which is good, other times words spread out evenly on the entire plot, like background noise, which is bad. Any measurement to quantify this?

  • $\begingroup$ the "clusteredness" of observations depends also on the chosen number of clusters. Have you tried varying that? $\endgroup$
    – utobi
    Commented Oct 11, 2022 at 14:22
  • $\begingroup$ @utobi The classes are premade, not a variable, so every word already has a label. I'm interested in how well my resulting embedding data is organized according to those premade classes. $\endgroup$
    – oliver.c
    Commented Oct 11, 2022 at 14:39
  • 1
    $\begingroup$ It looks like you have one partition (classes) and another partition (clusters). And want to measure how much the two agree. Right? Then you need some of external clustering criteria $\endgroup$
    – ttnphns
    Commented Oct 11, 2022 at 21:04
  • $\begingroup$ Given you have the "true labels" then @ttnphns link should cover you. $\endgroup$
    – usεr11852
    Commented Jun 2 at 23:52


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.