As the true user model is unknown beforehand, it is estimated
as the expectation over the posterior distribution of
the user model. Jin and Si [18] provides a computationally
efficient way of approximating the posterior distribution, but
still it is intractable for real recommender systems.