As Joseph points out, the real trick is dimensionality reduction. How you tune your dimensionality reduction algorithm and/or heuristic will have a significant impact on precision and recall (and performance).

Overall, however, LSA works very well for computing recommendations from textual data. Currently we’re applying PLSI to only the English Wikipedia corpus; however, it would probably work well on the IMDB dataset, too.

-ESer.org

Ultimately with LSA we are using a vector space model. No matter what way we derive the weights (pLSA, LSA, LDA) we map to that model.

So within the model we can use cosine to measure the relatedness of documents in the vector space.

With your example I can see you could find subtitles that where related. As for recommendations it really depends on what your movie data set is like. Whether it would make sense mapping a movie data item to a document in the vector space (where each vector is a subtitle document) and using cosine to find recommendations for subtitles.

I am a student and my thesis is on recommendation systems
i am working on a movie data set ( using svd )

i was wondering if we could use LSA on a subtitle data set
( if there is one ) in order to get recommendations

its just an idea, i see you dont have recommendations in
LSA’s applications so i just had to ask

thank you

My next release will be Latent Dirichlet allocation and this will be implemented in ruby. I’ll post here as soon as I’ve got the code up (and I’ll release the code before I write up the blog if you want to see it early).