-5

Currently, the default sort order for answers is by "Votes". This is simply upvotes - downvotes. It fails to consider the fact that some items simply have fewer votes because they are either newer answers or because they are currently so low down that they never get seen.

I would like to propose that we use projected score instead.

Here's a technical article on how to do that: http://www.evanmiller.org/how-not-to-sort-by-average-rating.html

And here's Randall Munroe (of XKCD) being less technical but more amusing on the subject: http://www.redditblog.com/2009/10/reddits-new-comment-sorting-system.html


Basically the way the equation (in the first link) works is that it looks at how many people have actually viewed something and sees how many of them upvoted/downvoted it. Based on that, it asks the question, if X people saw the item, what would its score be? It asks that question of every item and then sorts them according to that.

So it could see that one answer has 1000 views and 10 upvotes and another has 100 views and 5 upvotes, and it would reach the conclusion that when the second one has 1000 views, it'll have 50 upvotes, and therefore the second answer is better than the first.


Now that I'm thinking about this some more though, I'm realizing that while we're at it, we should factor in when votes were made. An answer that was great six years ago may no longer be great today. Still, I think using projected score would result in better sorting than the current system

  • Can you briefly explain how to do that, without linking to the article? – HDE 226868 Aug 30 '15 at 22:50
  • @HDE226868 - If you open the first link, it gives you the exact equation involved, along with an implementation of it in SQL. I know what the equation does and what the basic idea is behind it, but I don't really know how the person arrived at the specific equation. – ArtOfWarfare Aug 30 '15 at 22:55
  • @HDE226868 - I've expanded a bit on it in the second section of my post. – ArtOfWarfare Aug 30 '15 at 23:00
  • 1
    Note that we don't count answer views, just question views – Tim Aug 31 '15 at 10:54
  • 1
    @Tim - We would need to add "logged in answer views". We don't care about people who view answers without logging in, since they can't vote. There's a huge variety of ways you could choose to determine what is and isn't an answer view, but I would probably go with simply saying if the end of an answer is scrolled into view, the user has viewed it. – ArtOfWarfare Aug 31 '15 at 12:12
  • @ArtOfWarfare yes, I know we could but it doesn't currently happen, and I believe feature requests for it have been denied. – Tim Aug 31 '15 at 12:13
  • @Tim - Possibly they were requested with no greater reason described? Adding in that feature is a means to an end for my proposal, whereas it may have just been the end for other people proposing it. – ArtOfWarfare Aug 31 '15 at 12:15
6

In theory, I like this. In practice, I don't.

A linear prediction would be fine if an answer that starts off by getting six votes per 100 views continues to get six votes per 100 views. However, this is almost certainly not the case. For example, if an answer is posted and then the question hits the Hot Network Questions list, views from outside the site will skyrocket, and the post will be viewed many times by people who don't have accounts on the site and therefore will not vote. You're left with two predictions, both of which will probably disagree at or about that moment.

A similar phenomenon occurs when a question leaves the Hot Network Questions list. In fact, this will happen as the question jumps around and is displayed with different frequencies on different sites. It isn't easy to take that into account, and judging that the Hot Network Questions list accounts for some of the questions with the most votes, most views, and most answers - in other words, where this is needed most - predictions can be difficult.

Now, a non-linear model would be better - in fact, if curve fitting could be used to determine a best fit model - then things would be better. But, once again, where it matters most, the model fails.

This is just one manifestation of the same issue - formulas break down. There are a million[citation needed] different things that determine the votes on an answer: comparison with current answers, redundancy, edit(s), links to other questions. . . The list goes on. In the far future, new answers will have to be ranked by algorithms that take all this into account in order to come up with an accurate model.

You can't quantify human behavior in this issue without taking into account so many different factors, unfortunately. Who knows, maybe this will actually work. But I think that unless you can come up with a really detailed model, there will be inaccuracies, and that will be the problem.

  • 2
    Could we only count views from users which are logged in and so could vote? That way they don't get diluted by a flood of people viewing who don't vote because they can't. – ArtOfWarfare Aug 31 '15 at 1:04
  • @ArtOfWarfare That would be an excellent solution to the HNQ problem, yes. – HDE 226868 Aug 31 '15 at 12:25
  • 1
    Nearly two years later rereading this (someone just upvoted it so brought it back to attention), I feel like you're allowing perfect to be the enemy of better. Your argument is "we can't do it perfectly, so lets not try to make it better." It's an improvement, so it should be made. – ArtOfWarfare Apr 12 '17 at 15:40

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .