I strongly recommend prototyping this application with a traditional data mining tool like R. Simple linear regression is unlikely to work well here. My guess is that you may need to cascade several models in order to get good results. Some models might be a simple lookup or a non-linear transformation of a linear model, others might compute the logarithm of a correction factor. You may want to use a nearest neighbor model. In any case, you are going to require a number of iterations on this design to get it to work well.
On Wed, Oct 26, 2011 at 2:41 AM, Tommaso Teofili <[email protected]>wrote: > Hi Ted > > 2011/10/25 Ted Dunning <[email protected]> > > > Tommaso, > > > > Can you suggest a use case for this? What problem are you trying to > solve? > > > > I've used this to create a model to make predictions about how much one > "should" pay a product given existing training set of > <product-features,prices>. > > > > > > Are you suggesting something more advanced than just changing the > gradient > > on the existing stochastic gradient descent code? > > > > Have you seen the conjugate gradient code that was just committed? > > > > I didn't have a look at that code yet but I'll surely check both pieces and > see if / how I could manage to make it do the job. > Thank you very much. > Tommaso > > > > > > On Tue, Oct 25, 2011 at 12:05 AM, Tommaso Teofili < > > [email protected] > > > wrote: > > > > > Hi all, > > > recently I've been working with Octave [1] to implement the gradient > > > descent > > > algorithm for linear regression (with uni/multi features) and I wonder > if > > > such an implementation exists here in Mahout as Octave is good for > > > prototyping while I'd want to integrate this algorithm in a plain Java > > app. > > > Regards, > > > Tommaso > > > > > > [1] : http://www.gnu.org/software/octave/ > > > > > >
