"PCA with missing values"

keith

Does anyone have any ideas about how to compute principal components in cases where some of the (normalized) attributes in the data have missing values, other than assuming them all to be zero?  Do either of the PCA-related learners in RM have any facilities for dealing with missing values?

Thanks,
Keith

cantab

Stochastic gradient algorithms can compute PCA with missing data, simply by ignoring the unknown matrix entries.  This is an idea that has b=never really been properly published academically.  It came into prominence recently because it is very useful for collaborative filtering.  See

http://sifter.org/~simon/journal/20061211.html

Note that the real name of the author Simon Funk is Brandyn Webb.