Learning a representation from a set of vectors

Question

I'm currently dealing with the following problem: I have a set of feature vectors (real-valued) describing different instances of a common entity (such as an object or an event). Using these vectors, I would like to learn a common representation (a vector) for this entity (be it in the same vector space or a reduced one).

The most straightforward solution would be to use an arithmetic average. However, I was wondering if you could suggest some other solutions too?

Answer 1

You should also look at Principle Component Analysis (just google) and Sparse Dictionary Learning.

Answer 2

It's not entirely clear what the requirements are for the 'common representation' but you could have a look at Vector quantization.