0
$\begingroup$

I don't have a specific example for my problem and maybe this is trivial, but I want to know how to measure the influence of specific attributes (or dimensions) of a dataset for clustering, like there are ways to compute cluster validity.

As far as I know in PCA variance is used to determine which attributes can be unused but still no information is lost. Are there other units to measure the influence or importance of attribues to select the most promissing features? In other words: Is there a way to create a ranking or something to choose what attributes to use in a subset? Feel free to direct me to other sites or material if this question is to easy, but I couldn't find myself any desired answers.

$\endgroup$

1 Answer 1

-1
$\begingroup$

There is a simple approach (which is why this is not discussed a lot):

  1. Run the clustering algorithm
  2. Use the clusters as class labels
  3. Use an existing method to measure the importance of features for labeled data (e.g. decision tree, but also any other feature importance measure for labeled data)

Beware: clustering algorithms are usually very sensitive to scale. So it will not be very surprising to see a single feature dominate the result if the data was badly preprocessed.

$\endgroup$
2
  • $\begingroup$ thank you for your answer, I'm heading in the right direction now. Could you please name other examples of excisting methods to measure the importance of features or ways to find them myself? But the decision tree looks like something I was looking for. $\endgroup$
    – Urknecht
    Commented Jan 3, 2016 at 11:03
  • $\begingroup$ Please see "feature selection" on Wikipedia. $\endgroup$ Commented Jan 3, 2016 at 11:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.