A multilinear singular value decomposition (2000) [pdf] | Hacker News Reader

A multilinear singular value decomposition (2000) [pdf]

citeseer.ist.psu.edu

36 pointscinquemb5 comments

I can't download the paper. Can anyone provide a quick summary?

Abstract: We discuss a multilinear generalization of the singular value decomposition. There is a strong analogy between several properties of the matrix and the higher-order tensor decomposition; uniqueness, link with the matrix eigenvalue decomposition, first-order perturbation effects, etc., are analyzed. We investigate how tensor symmetries affect the decomposition and propose a multilinear generalization of the symmetric eigenvalue decomposition for pair-wise symmetric tensors.

Thanks andreyf.

I'm not familiar with most of those terms, but I'm guessing it's a big deal if it's got a decent number of upvotes. Do these findings have any benefits that a layperson might be able to understand (you can assume some programming knowledge if that's relevant)?

I was wondering about how someone would extend LSA techniques into tensor rank-k spaces where k > 2, so one could add as many datapoints/features collected other than terms and frequency of such (like parts of speech, sound clips, etc). I also want to use something like this for spacial filtering of eeg data.

Ideally, i would use the Jacobi-Davidson Method[0], but armadillo only has n x m partial singular value decomposition.

[0] http://netlib.org/utk/people/JackDongarra/etemplates/node340...

Attempt at a short summary:

A matrix M has an SVD U S V', where U and V have orthonormal columns and S is a diagonal matrix. How does this generalize to multidimensional arrays?

A third-order tensor T has a multilinear SVD of the form S x (U, V, W), where S is a core tensor and U, V, and W are again orthonormal matrices. In the matrix SVD, U transforms the columns of S, while V transforms its rows (hence it appears on the right of S). The product S x (U, V, W) is similar: U transforms the columns of the core tensor S, V its rows and W the mode-3 vectors of S.

The authors show that slices of S must be orthogonal under the inner product, e.g., <S(:,i,:), S(:,j,:)> = c * delta(i,j). In the case of a matrix SVD, this reduces to S being diagonal.

An example application is predicting how a user would rate a movie at a given time by approximating (the known entries of) a user x movie x time tensor with a truncated multilinear SVD.

Reply on news.ycombinator.com