0
$\begingroup$

I have a data of persons and for each person i have a list of feature : age, hair-color, skin-color, size, weight, ... I need to do hierarchical clustering on these feature, to group the feature which are "similar".

I try to do this, but I 've got as result a segmentation of persons: group of persons. But this is not what I hope.

Any idea please to help me?

$\endgroup$
0
$\begingroup$

Clustering should indeed return groups of items, in your case persons.

If you are interested in groups of similar features, and you have a small number of features I recommend the you'll compute the correlation among them. There are many correlation measures and you should choose the proper one according to your goals. Common correlations are mutual information, Pearson correlation and covariance. That will give you the correlation between pairs of features. Usually you can manually scan the list and see which features are close.

A more complicated tehcnique is to use your data set of person and build a dataset of presons' features. The items will be age, weight, etc. The distance function will be the correlation measure you choose. Then you run a clustering algorithm on this dataset and get groups of similar feature.

Please note that the typical use of clustering is your case will be to create groups of persons and not feature. If you'll describe more in the comments what you are looking for, maybe the community will be able to offer other methods.

$\endgroup$
  • $\begingroup$ Thank you Dan very much, I would like if there is a function like "pairs" in R which extract group of features , not only couple of feauture (2 features) for example 3 or more correlated feautures for exmple ( age, size, weight). Bests $\endgroup$ – Fish Apr 28 '16 at 8:54
  • $\begingroup$ You can use dbscan (also a clustering algorithm, has an R implementation). It will return connected components. In your case, features that are correlated directly or are correlated via some intermediate features. Since most correlation matrix are quite transitive, it will be very close to finding cliques of features. I suggest this method and not looking for exact solution to the clique since it is a NP-complete problem. $\endgroup$ – DaL May 1 '16 at 9:14

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.