Indeed not. But it *is* a case of the product architecture for recommendations I was nattering about.
The problem here is how to compute the (user x topic) x (item x topic)' product efficiently. This can be done pretty well with either hadoop or SQL. In Pig or native map-reduce, the trick is to group by the topic and then group by (user, item), summing the results as you go. If either user x topic or item x topic is small then a map-side join is good for the first group-by operation. If not, then doing two full-scale map-reduce operations is no big deal. You should consider how to weight different relevances, probably according to overall frequency in the corpus. On Mon, Jun 22, 2009 at 1:32 PM, Sean Owen <[email protected]> wrote: > I see. This almost is not a 'classic' recommendation problem. If you > have user-subject similarity, and subject-item similarity already, > then user-item similarity is probably just the product of the two? so > you can recommend items by ordering by similarity. >
