1
$\begingroup$

Is it possible to transform a document into a latent semantic space calculated with SVD without recomputing the SVD, or even doing a rank one update? I don't care about updating the $U$, $\Sigma$, or $V$ matrices, but I am interested in being able to calculate document similarity for an unseen document.

My linear algebra is very weak (apologies :), but given the SVD equation, $X = U \Sigma V^T$, my first thought was to calculate $U_{n+1} = X_{n+1} \left(\Sigma V^T\right)^{-1}$, but my solver fails because $\Sigma V^T$ is singular. Does this mean that I must do an update to get the new document into the same space?

$\endgroup$
0
$\begingroup$

I solved all my misconceptions by studying the Derivations section of the Wikipedia page on latent semantic analysis. In particular, a document in the latent semantic space is $\Sigma^{-1} U^T x$.

$\endgroup$

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.