This is interesting, especially considering that it favors low-data items, as opposed to both the confidence-interval-lower-bound and the notability adjustment factor, which penalize low-data items.

You can try to optimize it in an explore-vs-exploit framework, but there would be a lot of modeling parameters, and additional kinds of data will need to be considered. Specifically, a measure of how many of those who viewed the item bothered to vote at all. Some comments will not get any votes simply because they are not that interesting; so if you keep placing them on top hoping to learn more about them, you'll end up with very few total votes because you show people things they don't care about.