What should we do if clustering such as K-means is dominated by one or two variables in the list of used variables? Shall we leave the other variables?
$\begingroup$
$\endgroup$
2
-
2$\begingroup$ if "dominated" means there are large differences in order of magnitude: standardize the variables $\endgroup$– knbCommented Jul 30, 2018 at 7:25
-
1$\begingroup$ Yes, as said, you should consider standardizing but you don't have to standardize. It depends (see). In the end, when you are satisfied with theoretical considerations you've made and there is still the "dominance" in the sense that only one variable separates the clusters - then yes, you could drop other variables. $\endgroup$– ttnphnsCommented Jul 30, 2018 at 9:01
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
This is commonly caused by the variables having different scale.
How many people will just heuristically normalize the data. While this usually prevents one variable from dominating the result, whether or not this is meaningful is something you have to answer based on your data.
In my opinion, when you have scales of very different data types and scales (e.g., shoe size vs. income in $ vs. appartment size in ft²) then it is questionable if any k-mrans result on such data is meaningful...
answered Jul 30, 2018 at 11:46