LSI content based recommender system template

This workflow performs LSI text-mining content based recommendation.

We use SVD to capture latent semantics between items and words and to obtain low-dimensional representation of items. Latent Semantic Indexing (LSI) takes k greatest singular values and left and right singular vectors to obtain matrix 

A_k=U_k * S_k * V_k^T.

Items are represented as word-vectors in the original space, where each row in matrix A represents word-vector of particular item. Matrix U_k, on the other hand represents items in the k$ dimensional concept space.  We took our text-mining content-based recommender system template and transformed items from the word space to the k dimensional concept space. This technique captures latent semantics and merges similar words to the same concept.  

Afterwards, workflow calculates cosine similarity between items in the k dimensional concept space and recommends  top %{recommender_no} items. Also, %{recommendation_no} and %{reduction_k} are context defined macros which represent the required number of recommendations and reduction dimensionality constant k, respectively.