A system for generating recommendations which automatically optimizes over time without human intervention is disclosed. While known recommendation systems tended to be either attribute based or collaborative filtering based, the present system adjusts its internal parameters in response to the measured performance of the recommendations as determined by user behavior. In one embodiment, the user provides feedback in the form of ratings for the recommendations. This feedback is processed and then utilized to adjust the internal parameters of the system. Over time, the system settles on optimal parameter values.