post-clustering feature selection high dimension data

Question

I would like to know if there are very popular methods for feature selection following clustering with k-means or HAC.

More precisely, I used these methods on genomic data to sort a hundred patients based on about 2,500 genes. Now I would like to select the genes wich are the most important for patient sorting.

Does anybody have any idea ?

Thanks !

Answer 1

I don't think you need anything specific to clustering. Just identify which features best separate your data.

For example, you can train a decision tree or random forest to separate your cluster from the remainder of the data. This will give you important attributes.

Or you could use a Fisher's linear discriminant and check the direction of the resulting vector if a linear separation works.