|
Hi, I'm working with a really sparse matrix (way less than 1% sparse) trying to infer missing values with Matrix Factorization. Is there any theory on "accuracy" with really sparse data or any other techniques that purport to perform on this type of matrix? thanks in advance, Jonny |
I never dealt with the problem myself but may be you want to check Emmanuel Candes and Yi Ma's work on low rank matrix factorization.
http://www-stat.stanford.edu/~candes/papers/MatrixOracle.pdf
http://perception.csl.illinois.edu/matrix-rank/Files/RPCA_JACM.pdf
I would imagine that all the work for the NetFlix competition (inferring the missing values of a user, value => score matrix) are relevant here.