Optimal Scoring for Unsupervised Learning