KDD Cup 11 - SVD based Recommender

[SJ <sa...@gmail.com>]

Hello,
I have tried using Mahout (taste- non distributed version) to test
out-of-the-box performance on KDD cup, and for track 1 found an RMSE of 29.09,
which is certainly not bad (for out-of-the-box Item-based recommender,
PearsonCorrelationSim)

However, I would like to use the SVD-based recommender, for further evaluation.
There are ALSWRFactorizer and ExpectationMaximizationSVDFactorizer factorizers
available, but both require a large number of parameters and knobs to tune
(numIterations, numHiddenFeatures, lambda, preventOverfitting,randomNoise etc. )

If anyone has worked with these recommenders before, could you please help me
out, how to go about tuning these params, for my dataset. And wether it would
even be a good idea to work with these recommenders for large datasets (650k
items, 1m users) I would expect SVD factorizations to help while dealing with
sparse data, but am not familiar with the Taste implementations, whether they
scale (memory, time required) for these large datasets.

Thanks for your help.

Regards,
SJ