An Optimized Interaction Strategy for Bayesian Relevance Feedback