Comments on: Why Does Latent Semantic Analysis Work? http://thenoisychannel.com/2008/11/01/why-does-latent-semantic-analysis-work/ Sun, 12 Feb 2012 20:45:02 +0000 hourly 1 http://wordpress.org/?v=3.3.1 By: Daniel Tunkelang http://thenoisychannel.com/2008/11/01/why-does-latent-semantic-analysis-work/comment-page-1/#comment-693 Daniel Tunkelang Tue, 04 Nov 2008 03:54:25 +0000 http://thenoisychannel.com/?p=729#comment-693 Very cool. Will put that on my reading queue. By the way, does anyone know if the patent on LSA (4,839,853) has expired? According to Wikipedia, "For applications that were pending on and for patents that were still in force on June 8, 1995, the patent term is either 17 years from the issue date or 20 years from the filing date of the earliest US application to which priority is claimed (excluding provisional applications), the longer term applying." I think that means that the patent expired in September. Very cool. Will put that on my reading queue.

By the way, does anyone know if the patent on LSA (4,839,853) has expired? According to Wikipedia, “For applications that were pending on and for patents that were still in force on June 8, 1995, the patent term is either 17 years from the issue date or 20 years from the filing date of the earliest US application to which priority is claimed (excluding provisional applications), the longer term applying.” I think that means that the patent expired in September.

]]> By: Bob Carpenter http://thenoisychannel.com/2008/11/01/why-does-latent-semantic-analysis-work/comment-page-1/#comment-692 Bob Carpenter Mon, 03 Nov 2008 23:04:53 +0000 http://thenoisychannel.com/?p=729#comment-692 There's a really nice description of the large space of SVD-like factor analyses and the relations among them in Aleks Jakulin and Wray Buntine's paper <a>Discrete Component Analysis</a>. It talks specifically about how the various SVD-like factorings take variance and sparsity into account. You can find implementations of SVD for complete matrices in packages like R and Matlab, but the folks who used it for Netflix invented a missing-data variant based on stochastic gradient descent that's so neat I implemented it in LingPipe (see our <a href="http://alias-i.com/lingpipe/demos/tutorial/svd/read-me.html" rel="nofollow">singular value decomposition tutorial</a>, which contains lots of references and an implementation of the demos from the original latent semantic indexing papers). There’s a really nice description of the large space of SVD-like factor analyses and the relations among them in Aleks Jakulin and Wray Buntine’s paper Discrete Component Analysis. It talks specifically about how the various SVD-like factorings take variance and sparsity into account.

You can find implementations of SVD for complete matrices in packages like R and Matlab, but the folks who used it for Netflix invented a missing-data variant based on stochastic gradient descent that’s so neat I implemented it in LingPipe (see our singular value decomposition tutorial, which contains lots of references and an implementation of the demos from the original latent semantic indexing papers).

]]>