Rejection sampling is a useful technique for performing supervised learning on training sets too large to be learned in their entirety. FilterBoost is a recent extension to AdaBoost which uses rejection sampling in an online learning framework and has been shown to work for automatic tagging of music. In this paper we improve on FilterBoost by adding Metropolis-Hastings sampling, thus allowing the algorithm to focus on hard-to-classify examples. We describe how our knowledge of artist-level similarity can be used effectively in a Metropolis-Hastings framework and demonstrate a significant increase in classification accuracy over standard FilterBoost.