Is it possible to modify the two Factorizer implementations to generate and save the singular values? There is an alternate SVD-based recommender algorithm that does a nice job with "anonymous" users. It needs either the singular values, or the inverse of the singular values baked into the left matrix.
On Thu, Sep 15, 2011 at 10:46 PM, Sebastian Schelter <[email protected]> wrote: > It did not mean incremental SVD, it just means that refresh() can be > called on an SVDRecommender and it will recompute its decomposition: > > https://issues.apache.org/jira/browse/MAHOUT-640 > > --sebastian > > 2011/9/16 Lance Norskog <[email protected]>: > > For the 0.5 release, these are mentioned: > > > > * Incremental SVD Implementation > > * SVDRecommender enhancements > > > > This seems to be about adding a refreshable SVD-based recommender. > > The Factorizers do not contain an incremental SVD implementation, like > this: > > > > Fast Low-Rank Modifications of the Thin Singular Value Decomposition > > http://www.merl.com/publications/TR2006-059/ > > > > Was this the intention? > > > > -- > > Lance Norskog > > [email protected] > > > -- Lance Norskog [email protected]
