# Classifier on top of LDA topic vectors?

I have training data in form of pair of documents with an associated label - {doc1, doc2, label}. Label is defined as function of pair of documents.

Now I want to build a model which can predict the label given two new documents.

I want to try different representation of document (instead of common ones say TF-IDF). Can I use vectors (topic distribution) from LDA as features for a classifier?

• Instead of concatenation, I'd lean towards learning an inner product; have the input to sigmoid function be $x_1^\dagger M x_2$, for a PSD M, where the parameter to be learned is now the matrix $M$. But go ahead try it your way, see what you get. – Emre Sep 11 '16 at 21:19