Comments on: SVD Recommendation System in Ruby http://www.igvita.com/2007/01/15/svd-recommendation-system-in-ruby/ A goal is a dream with a deadline. Fri, 12 Mar 2010 22:38:38 +0000 http://wordpress.org/?v=2.7.1 hourly 1 By: InVisible Blog » links for 2010-03-05 http://www.igvita.com/2007/01/15/svd-recommendation-system-in-ruby/comment-page-2/#comment-232593 InVisible Blog » links for 2010-03-05 Fri, 05 Mar 2010 16:02:59 +0000 http://www.igvita.com/blog/2007/01/15/svd-recommendation-system-in-ruby/#comment-232593 [...] – without algebra (tags: ruby datamining lsi algorithms tagging programming data python) SVD Recommendation System in Ruby – igvita.com And a detour on SVD in Ruby (tags: ruby programming rails datamining learning rubyonrails [...] [...] – without algebra (tags: ruby datamining lsi algorithms tagging programming data python) SVD Recommendation System in Ruby – igvita.com And a detour on SVD in Ruby (tags: ruby programming rails datamining learning rubyonrails [...]

]]> By: Yaqing http://www.igvita.com/2007/01/15/svd-recommendation-system-in-ruby/comment-page-2/#comment-232202 Yaqing Tue, 02 Mar 2010 17:35:00 +0000 http://www.igvita.com/blog/2007/01/15/svd-recommendation-system-in-ruby/#comment-232202 Hey, nice post,, thx so much. :) btw,,, i am not very good at mathematics and matrix calculations,,, can u give the formular for Embed new item also?... Thx very much... Hey, nice post,, thx so much. :)

btw,,, i am not very good at mathematics and matrix calculations,,, can u give the formular for Embed new item also?… Thx very much…

]]> By: tjumicy http://www.igvita.com/2007/01/15/svd-recommendation-system-in-ruby/comment-page-2/#comment-224330 tjumicy Thu, 26 Nov 2009 07:03:51 +0000 http://www.igvita.com/blog/2007/01/15/svd-recommendation-system-in-ruby/#comment-224330 As egarcia commented on 06.05.07, it is really confused me until I manually put the matrix to SVDLIBC and get the real U、S、V I think the image ( http://farm1.static.flickr.com/133/358494623_db22603640_o.png containing U S V.transpose) should be U S V. The title in the image "V.transpose" should be "V". :-) As egarcia commented on 06.05.07, it is really confused me until I manually put the matrix to SVDLIBC and get the real U、S、V

I think the image ( http://farm1.static.flickr.com/133/358494623_db22603640_o.png containing U S V.transpose) should be U S V. The title in the image “V.transpose” should be “V”. :-)

]]> By: tjumicy http://www.igvita.com/2007/01/15/svd-recommendation-system-in-ruby/comment-page-2/#comment-224329 tjumicy Thu, 26 Nov 2009 06:34:26 +0000 http://www.igvita.com/blog/2007/01/15/svd-recommendation-system-in-ruby/#comment-224329 Nice post. Thanks a lot. During reading, I noticed there maybe a typo at the first paragraph above the MxN matrix table image: <blockquote> Thus, with M users and N products, we are looking at an MxN matrix. </blockquote> should be <blockquote> Thus, with M products and N users, we are looking at an MxN matrix. </blockquote> We can see M = 6, N = 4, U is (MxM) indicating products, V is (NxN) indicating users. So the table MxN should be 6x4 with 6 products and 4 users which I mentioned above. Hope I'm not wrong and it can be helpful above to fix this typo. Nice post. Thanks a lot.

During reading, I noticed there maybe a typo at the first paragraph above the MxN matrix table image:

Thus, with M users and N products, we are looking at an MxN matrix.

should be

Thus, with M products and N users, we are looking at an MxN matrix.

We can see M = 6, N = 4, U is (MxM) indicating products, V is (NxN) indicating users. So the table MxN should be 6×4 with 6 products and 4 users which I mentioned above.

Hope I’m not wrong and it can be helpful above to fix this typo.

]]> By: Dharmendra http://www.igvita.com/2007/01/15/svd-recommendation-system-in-ruby/comment-page-2/#comment-223372 Dharmendra Sun, 15 Nov 2009 17:25:56 +0000 http://www.igvita.com/blog/2007/01/15/svd-recommendation-system-in-ruby/#comment-223372 I think you shouldn't use S^-1 in projecting Bob (new vector) onto the latent space. The columns of U are the orthonormal "users" which form the bases for representation of real users. If you use U*S as the bases, then it becomes orthogonal bases, not orthonormal. Two users' equivalent latent representation would be U*S*v1 and U*S*v2 where v1 and v2 are the column vectors of the matrix V.transpose. Then similarity in the latent space would be (U*S*v1).transpose*(U*S*v2)/(their norms) = v1t*S^2*v2/(sqrt(v1t*S^2*v1)*sqrt(v2t*S^2*v2)) which is not equal to v1t*v2/(sqrt(v1t*v1)*sqrt(v2t*v2)) as used in the example. I think you shouldn’t use S^-1 in projecting Bob (new vector) onto the latent space.

The columns of U are the orthonormal “users” which form the bases for representation of real users. If you use U*S as the bases, then it becomes orthogonal bases, not orthonormal.

Two users’ equivalent latent representation would be U*S*v1 and U*S*v2 where v1 and v2 are the column vectors of the matrix V.transpose. Then similarity in the latent space would be
(U*S*v1).transpose*(U*S*v2)/(their norms)
= v1t*S^2*v2/(sqrt(v1t*S^2*v1)*sqrt(v2t*S^2*v2))

which is not equal to

v1t*v2/(sqrt(v1t*v1)*sqrt(v2t*v2)) as used in the example.

]]> By: Ilya Grigorik http://www.igvita.com/2007/01/15/svd-recommendation-system-in-ruby/comment-page-2/#comment-223311 Ilya Grigorik Sat, 14 Nov 2009 17:19:50 +0000 http://www.igvita.com/blog/2007/01/15/svd-recommendation-system-in-ruby/#comment-223311 @Musi: unfortunately don't have a Java equivalent. However, I don't think JRuby would be out of the question -- in fact, it might even be a better choice for the heavy number crunching involved in SVD's. @Musi: unfortunately don’t have a Java equivalent. However, I don’t think JRuby would be out of the question — in fact, it might even be a better choice for the heavy number crunching involved in SVD’s.

]]> By: musi http://www.igvita.com/2007/01/15/svd-recommendation-system-in-ruby/comment-page-2/#comment-223182 musi Thu, 12 Nov 2009 16:48:19 +0000 http://www.igvita.com/blog/2007/01/15/svd-recommendation-system-in-ruby/#comment-223182 I applied SVD over A sparse matrix using java, the results are not as expected . I am using JAMA package and working on recommender systems? Do you have any code in Java? I applied SVD over A sparse matrix using java, the results are not as expected . I am using JAMA package and working on recommender systems?
Do you have any code in Java?

]]> By: Jesse Storimer http://www.igvita.com/2007/01/15/svd-recommendation-system-in-ruby/comment-page-2/#comment-220841 Jesse Storimer Sun, 18 Oct 2009 12:32:17 +0000 http://www.igvita.com/blog/2007/01/15/svd-recommendation-system-in-ruby/#comment-220841 Fantastic post! Really enjoyed it. I had some trouble installing rb-gsl on Snow Leopard. Here is how I did it: http://www.jstorimer.com/blogs/my-blog/1284032-installing-rb-gsl-on-snow-leopard Also, the rb-gsl example seems to have fallen out of date. These are the changes that worked for me: Lines 22 and 23: remove the .transpose call Line 33: change to user_sim[count] = (GSL::Vector[bobEmbed[0],bobEmbed[1]].dot(x)) / (x.norm * bobEmbed.norm) http://gist.github.com/212661 Fantastic post! Really enjoyed it.

I had some trouble installing rb-gsl on Snow Leopard. Here is how I did it: http://www.jstorimer.com/blogs/my-blog/1284032-installing-rb-gsl-on-snow-leopard

Also, the rb-gsl example seems to have fallen out of date. These are the changes that worked for me:

Lines 22 and 23: remove the .transpose call
Line 33: change to
user_sim[count] = (GSL::Vector[bobEmbed[0],bobEmbed[1]].dot(x)) / (x.norm * bobEmbed.norm)

http://gist.github.com/212661

]]> By: ekzept http://www.igvita.com/2007/01/15/svd-recommendation-system-in-ruby/comment-page-2/#comment-219894 ekzept Sat, 10 Oct 2009 20:19:27 +0000 http://www.igvita.com/blog/2007/01/15/svd-recommendation-system-in-ruby/#comment-219894 There is another interesting aspect to all this, and that is the coding plan equivocates a person not liking a property or season of a TV show with not knowing whether or not a person likes a property or season of a TV show. Properly speaking, those are different states of knowledge and should be accounted for in the model. There is another interesting aspect to all this, and that is the coding plan equivocates a person not liking a property or season of a TV show with not knowing whether or not a person likes a property or season of a TV show. Properly speaking, those are different states of knowledge and should be accounted for in the model.

]]> By: ekzept http://www.igvita.com/2007/01/15/svd-recommendation-system-in-ruby/comment-page-2/#comment-219892 ekzept Sat, 10 Oct 2009 20:14:00 +0000 http://www.igvita.com/blog/2007/01/15/svd-recommendation-system-in-ruby/#comment-219892 Incidently, limiting the display to two dimensions -- corresponding to the top two singular vectors -- is somewhat arbitrary and is done for easier display and understanding. If the corresponding singular values are significant, in general, you'd want to extend the description to more dimensions, although a Euclidean distance (L2 norm) may still be appropriate. If there is a lot of data, you might want to partitition the space somehow, perhaps using a Dirichilet tesselation or quantiles separating points. Incidently, limiting the display to two dimensions — corresponding to the top two singular vectors — is somewhat arbitrary and is done for easier display and understanding. If the corresponding singular values are significant, in general, you’d want to extend the description to more dimensions, although a Euclidean distance (L2 norm) may still be appropriate. If there is a lot of data, you might want to partitition the space somehow, perhaps using a Dirichilet tesselation or quantiles separating points.

]]>