You might want to weigh less controversial (as in abs(upvotes - downvotes) higher. This would be somewhat like Effect Size in science. The Bayesian approach wo

You might want to weigh less controversial (as in abs(upvotes - downvotes) higher. This would be somewhat like Effect Size in science. The Bayesian approach would be to assume the true vote distribution is binomial and use a beta prior (possibly with Jeffrey's degenerate bimodal prior). Then as the total number of votes increases the posterior distribution tightens. Ranking score is prob(score>0).