# Is it possible to select features from completely unlabeled data?

I have seen many examples of using semi-supervised learning to reduce the the number of features in a data set, but I am wondering if it is possible to somehow reduce features with purely unlabeled data.

Trivially, we can remove features that are the same value for every instance or that are very redundant in data, as these obviously cannot help us when generating a model. However, might there be some other, more non-trivial manners we may remove features from unlabeled data?

Intuition tells me that this is not possible as you can only know what a feature does/measure its worth if you can see its effect on data, but perhaps there was some caveat I have not considered. I have done some searching on here and online but all seem to include positive examples. (I have also found papers with abstracts that seem to indicate this might be possible, but the papers themselves cost \$ which I would prefer not to spend).

Any answers/comments (even if they have links to papers) would be much appreciated, even if this is not know to be possible.

• What is the aim of your analysis? Why you need to select the features? What for? – Tim Aug 15 '18 at 19:33
• @Tim am doing a project that involves clustering and I have many features currently, which I would like to boil down as much as possible while still getting good results. – Alerra Aug 15 '18 at 19:36
• I'm sure you've heard about autoencoders. Why don't they fit your needs? – Aksakal Aug 15 '18 at 19:39
• You might want to consider subspace clustering algorithms, which are meant to work with high-dimensional data. – sebp Aug 15 '18 at 19:42
• @Aksakal I have done some investigation on autoencoders, and while they do look very useful for my needs and I will probably incorporate them into my project in some way, I am looking for as many possible methods as I can so that way I can get the best model for my needs. Appreciate the input! – Alerra Aug 15 '18 at 20:17