I have performed hierarchical clustering on a data set with 186 participants and 94 variables for each participant.

What I want to know is if there is a way to see which features are "driving" my clustering as such? For example is it age, gender etc. that is splitting my data set into the clusters.

Due to the number of variables I don't have time to go through each cluster member and check which variables allow it to belong to its cluster.

Hope this makes sense. C

  • $\begingroup$ Among internal clustering criteria there is Ratkowski-Lance one which is handy in that it can be easily computed for each individual feature as well (in addition to the overall all-feature value, the average). So, you can see then which features are more and which are less important in determining the cluster partition of your data. $\endgroup$
    – ttnphns
    Jan 4, 2019 at 9:23

1 Answer 1


You can use any of the correlation or feature importance measures.

If you didn't prepare your data well, it's most likely the one of highest range anyway...

For example random forests can be useful at identifying the relevant features. But even a single decision tree can be helpful already.

For correlations, you better use Spearman etc., and need to experiment with the ordering of clusters, as a cluster could be between others in any order. I'd give each cluster a rank after sorting their means, and then try to correlate the element ranks with the cluster ranks.


Your Answer

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

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