The users of SVD on Mahout are more interested in their own application level metrics than the metrics associated with the decomposition itself. Moreover, at the scale people have been working with, getting any decomposition at all is refreshing.
The examples that you gave in your thread involved walking *way* down the spectrum to the smaller singular values. That is absolutely not the interest with most Mahout users because that would involve fairly massive over-fitting. 2011/12/19 Radim Rehurek <[email protected]> > No problem. > > If you decide to run some SSVD accuracy experiments in the future (don't > your users ask for that?), please include me in cc -- just in case I miss > the post on this mailing list. >
