The idea behind latent semantic analysis is factorizing Mdw so as to extract a set of latent variables (this means that we can assume their existence but they cannot be observed directly) that work as connectors between the document and terms. As discussed in Chapter 11, Introduction to Recommendation Systems, a very common decomposition method is SVD:
However, we're not interested in a full decomposition; we are interested only in the subspace defined by the top k singular values:
This approximation has the reputation ...
