8
$\begingroup$

I have data that represent some aspect of human behavior. I want to cluster it (unsupervised) into behavioral profiles of some sort. now, some of my variables are categorical (with 2 or more categories), and some are continuous (most are percentages). A few variables are even more complex in that one category has further continuous and the other one has no such additional data.

My question is about how to go about categorize this data. What are the (common?) approaches dealing with it?

I don't need code or anything, but rather some references or directions that will help me further understand how to deal with this challenge.

If you know of R functions that facilitate such analysis, that would be great, but it's not necessary.

thanks.

$\endgroup$
  • 2
    $\begingroup$ Gower similarity measure can take simultaneously continuous, ordinal, binary, nominal data. You can use such clustering methods as hierarchical or medoid, to analyse the proximity matrix. Few other clustering methods (e.g. TwoStep cluster) can take continuous and nominal variables at once. $\endgroup$ – ttnphns May 25 '14 at 12:14
  • $\begingroup$ As for percentages or counts, sometimes special chi-square measures are computed for them, and sometimes usual euclidean distance, as for continiuos data, is used. $\endgroup$ – ttnphns May 25 '14 at 12:16
  • 4
    $\begingroup$ All in all, clustering mixed-type data is a tricky thing and might be for an experienced data analyst only, perhaps. On the other hand, clustering of such data is often not a good idea at all, because there are issues of standardization, interpretation and features contribution analysis. $\endgroup$ – ttnphns May 25 '14 at 12:21
1
$\begingroup$
  1. Spend lots of time on understanding similarity on your data.
  2. Formalize your notion of similarity in a specialized similarity measure, designed for your particular data set (you will likely not be able to use an out-of-the-box similarity).
  3. Use a clustering algorithm that can use arbitrary similarites, such as hierarchical clustering, DBSCAN, affinity propagation, or spectral clustering.
$\endgroup$
1
$\begingroup$

See https://cran.r-project.org/web/packages/ClustOfVar for the R package ClustOfVar. It appears to implement some of the best available clustering methods for mixtures of variable types.

$\endgroup$

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.