I'm running flat clustering algorithm on my dataset that contains numeric (not categorical) data.

Is there a method that can give me interpretation of clusters, and emphasize what are the most important variables and their range of values in cluster?

Something like - This cluster contains mostly elements for which Property XY has small values, and Property YZ has great values?

Or do I need post processing, like count elements in cluster that have value above certain threshold and something similar?


  • $\begingroup$ The question is really not clear. Presumably, by "This parameter" you mean 'this cluster'. What could "Parameter XY" etc mean? Do you mean, 'the mean vector'? What do you mean by "the most important parameters"? How can 1 parameter be more important than another? Etc. $\endgroup$ – gung - Reinstate Monica Oct 13 '14 at 15:57
  • $\begingroup$ @gung I guess "parameter" here from eugen means "variable" in more usual statistical parlance. People in several sciences use "parameter" as an all-purpose synonym for property or variable; not only does that seem odd to people saturated in statistics, it contradicts the meaning at least some of us learned in elementary mathematics, but there you go. $\endgroup$ – Nick Cox Oct 13 '14 at 17:18
  • $\begingroup$ @Nick Cox thank you for clarifying that. Yes, I have meant variable when I said parameter. $\endgroup$ – Kobe-Wan Kenobi Oct 13 '14 at 18:55

Depending on how many variables you have, you could plot them pairwise, color-coding by cluster, and eyeball it. You'd be looking for patterns of where the colors (your clusters) fall in each graph.

Beyond that, a wild thought is that you could take your data and create a categorical variable based on your clusters, then use a (classification) decision tree with the clusters as its target.

How you create this categorical variable depends on your software, but the idea is that you don't want to literally use cluster number as if it was a continuous variable. It needs to be categorical.

Once you've added the cluster identifier to each data point, you can use that as a target (the dependent variable) for in a decision tree. If you're using R, I found a good blog posting that should give you a start with several R packages: tree, rpart, party, partykit, maptree, evtree, etc. Harder to interpret but more flexible would be MARS (implemented in R's earth package).

| cite | improve this answer | |
  • $\begingroup$ Thank you for your answer Sir, but can you explain your approach a bit more, perhaps with some example? Your second approach with decision trees sounds better to me, since I'm trying to automate the procedure in some way... $\endgroup$ – Kobe-Wan Kenobi Oct 13 '14 at 15:32
  • $\begingroup$ OK, I'll edit a bit, though I can't think of any simple examples. $\endgroup$ – Wayne Oct 13 '14 at 15:34

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.