On 10.05.2012 17:33, 冯伟 wrote: > I want to look at the distribution implementation of matrix factorization > in Mahout Recommender System. Before I start from > org.apache.mahout.cf.taste.hadoop.als.RecommenderJob，is there any papers / > technical materials for reference? It seems that the parameters are learned > by ALS. Then is there a stochastic gradient descent implementation? I know > GraphLab of CMU for quite a while since KDDCup 2011，is there any comparison > between GraphLab's collaborative filtering lib and Mahout's?
Mahout's implementation is based on the following papers: Large-scale Parallel Collaborative Filtering for the Netflix Prize http://www.hpl.hp.com/personal/Robert_Schreiber/papers/2008%20AAIM%20Netflix/netflix_aaim08(submitted).pdf Collaborative Filtering for Implicit Feedback Datasets http://research.yahoo.com/pub/2433 There is a comparison in the original Graphlab paper which is a little biased IMHO because it uses an initial hacky version of the ALS implementation and the experiment is run on a really small dataset. I still think that Mahout's implementation will be something like 20x slower than GraphLab mainly due to Hadoop's inability to efficiently run iterative computations. Mahout only has a non-distributed SGD implementation of matrix factorization. --sebastian
