Recommender system - mean average precision metric optimization

Question

I'm trying to build a recommender system that i must use off-line, on fixed training and test sets, and that i must evaluate with the precision metric. Each user gave some ratings (1-5 values), so I don't have null rows in the user ratings matrix. I must recommed items for the users in the test set, all of them are also present in the training set, and i must optimize the algorithm for the precision metric (in particular MAP@5, mean average precision). Someone, on the base of his experience, can give me an advice on the algorithm that I should use? And maybe also on how i should implement it, if I should use particular tools/libraries or a particular programming language.

Answer 1

This might come a bit late, but I'm doing a similar task (if I understood yours correctly) with Mahout and so far, so good. Concerning MAP (this is actually why I stumbled upon your question): as far as I could see, this is not a out-of-the-box feature in Mahout. You might want to have a look at Myrrix (e.g. this), I saw several posts that referred to this. BUT I must warn you that I didn't try it myself...