One approach is to recommend content items as well as items and fold these recommendations together. Content-based recommendations are then done by retrieving items with weighted combinations of the recommended content values. Retrieval of items from content should be roughly in descending popularity order.
One way to fold the results is simply to take the content recommendations in cold start situations, the item recommendations for hot items and the union for intermediate situations. A slightly more principled way to combine is to compute a score for each item in the union of content and item-based recommendations using something like an SGD model. The inputs to the model are the content score (if any), the item score (if any) and a variety of features that help the model decide how to transition between the two recommendation models. Another approach is to recommend from a combined set of items and content values. To generate the final result, dereference the content values, but pass the items through verbatim. Since you usually need a retrieval step to decorate the items with additional data, this dereference of content and item id's is often free. On Fri, Feb 25, 2011 at 10:47 AM, Marko Ciric <[email protected]> wrote: > Hi guys, > > I'm currently using a generic item-based recommender as a content-based > recommender by customizing the item similarity to compare item features. > Basically, I want to solve a cold-start problem I have with my small data > set. I'm wondering now, is there a way I can use Taste existing recommender > evaluators to evaluate my content-based recommender. Any hints? > > -- > -- > Marko Ćirić > [email protected] >
