0
$\begingroup$

I have some data that I want to get the important variables from. I want to use random forest to get this information. The problem is that the data does not have labels. From what I understand random forest is a supervised learning algorithm, but one can generate synthetic data to help with the prediction.

Can random forest look for an underlying pattern in the data? If so how would I do this? Essentially can I get the important variables from this unlabeled data and how?

How does this compare to clustering?

Some source:

https://www.stat.berkeley.edu/~breiman/RandomForests/cc_home.htm#unsup

$\endgroup$
  • $\begingroup$ cforest doesn't have functions that readily provide this suggestion by Breiman. It shouldn't be too hard to code it yourself and just try it on your data to see how whether it gives results that are useful for you. $\endgroup$ – Achim Zeileis Jan 4 '16 at 9:14
2
$\begingroup$

Random forest are ensembles of decision trees.

By leaving out parts pf the variables and data they tend to abstract better, and overfit less. But if you don't have labels, overfitting is not much of a concern.

However, you cannot even build a single decision tree without labels.

The reason is probably unfixable: to decide upon a split (i.e. to creat a node in the tree) you need labels to evaluate the quality of your options...

So no, I don't you can use it on unlabeled data.

You could use a clustering algorithm to creat labels for your data.

But that will likely be a self-fulfilling prophecy. Clustering algorithms are sensitive to scale, so of course the features with the largest scale/variance have the largest importance.

$\endgroup$

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.