Solving the problem that the topmost comments get all upvotes
This assumes a discussion system that sorts comments by number of upvotes.
[[Update: As of August 2015, I have partly implemented the approach in the first example below (but not the second example). ]]
A fairly prevalent problem in discussions online: Some early
posted comments are fairly useful, and get fairly many upvotes,
but no downvotes (since they are fairly useful). Later on, some really
interesting comments are posted — but at that time, the early comments have already
accumulated many upvotes.
The early comments are therefore shown first, and continue to
gather more and more upvotes, because they receive all attention.
The really interesting comments, however, remain forgotten somewhere below,
because too few people take the time to scroll down, find them and read them.
This should solve the above-mentioned problem:
The computer counts how many people have read each comment, and takes this into
account, when it sorts all comments. — So, we don't sort by upvotes only, but
by upvotes divided by attention (that is, how many people have read a
The fairly useful comments mentioned above got many upvotes, but much more
attention than upvotes. Therefore the sorting algorithm would move them downwards,
and the really interesting comments would take their place
— because the really interesting comments have been upvoted very much,
compared to how many people have read them.
The result should be that really interesting comments surface to the top of
the page, even if they are posted much later than the early comments.
Please note that dividing by downvotes won't work,
because early posted fairly useful comments get virtually no downvotes, only upvotes.
Promoting new comments, in some manner, until they have received a few ratings, won't work well, I think. The result would be top of the page covered with lots of new mediocre comments?
How does the computer know which comments you've read? Here follows two examples
of how the computer could deduce which comments you've read.
First example. In the picture below, if you upvote the
comment outlined in orange, the computer would assume you've read all
comments leading up to that comment, plus any siblings placed before it.
Those comments are outlined in blue.
Second example. The computer could assume that you read comments that are shown on screen.
Have a look at this demo page — [[sorry the demo is broken nowadays (year 2015) because of lots of changes in how the software works]]. In the demo, a red square indicates that the computer thinks you have not yet read the comment. A blue square indicates that the computer thinks you've read it. Look at the demo page for a while, and watch the squares change color to blue.
In the two examples above: When you upvote a comment, the computer thinks that the other comments you have read (the blue ones) but did not upvote, are not terribly interesting.
This problem elsewhere, and solved?
Search engines have a similar problem. They estimate the usefulness of a search result link, by counting how many people click the link. However people tend to click the topmost links only. Here are some articles about how search engines approach this problem (I have not yet read them in whole):
- Cascade Model (one of the original click models)
- Dynamic bayesian network model (a more generalizable Cascade Model).
As far as I can tell, example 1 above reminds of the simplified model described
in section 5. A SIMPLIFIED MODEL.
- DBN model with scroll and hover interactions (reminds of example 2 above)
(Thanks to whathappenedto at Hacker News for posting these links.)
Is it a good idea to take into account how many people have read a comment? And not consider only upvotes and downvotes. Are there problems I've overlooked?
- 61 replies
- AAnonymous??2013-03-24 01:50Z
I was an early poster on a political discussion and got a single down vote. That pushed my comment down to the lowest imbedded level. So unless you were showing level 7 comments it didn't even show up. Totally unfair that one disagreeing thumbs down buried my comments.
- AAnonymous??2013-03-24 01:54Z
This is unequivocally the most interesting comment in the world, and yet here it lies at the bottom...hmmm, call me a skeptic.
- AAnonymous??2013-03-24 03:44Z
Sometimes "highly" disagreed with comments can be interesting to read as well (to see what kind of opinion many people find untrue), so perhaps both could be pushed to the top -- comments with a high upvote to view ratio, and comments with a high downvote to view ratio.
In your case, your comment might remain high with only a single down vote, until either (1) time passes and more people view it without voting on it, or (2) up and down votes even out.
- Uuser_145 (KajMagnus)2013-03-25 10:47Z
I think it'd be interesting help people find strongly controversial comments. I'm not sure about comments that virtually everyone votes down though — in general, I'd guess such comments tend to waste peoples' time (although there are exceptions I suppose).
Finding highly upvoted comments that differs from the current reader's point of view would be really interesting I think :-) But how does the computer know what the current reader thinks, and what a comment "thinks"?
- Uuser_145 (KajMagnus)2013-03-24 09:42Z
That can be fixed, and is already fixed on this website :-)
Details: I think the approach suggested on this page should be combined with algorithms that take into account that one doesn't know how interesting a comment is, after only a few votes. Only after many votes, it's possible to make a good estimate.
One way to do this, should be to sort by the lower bound of a confidence interval of the true score of the comment. — This is actually how the discussion system on this website already works :-) So the problem you're referring to, should have been largely solved already, on sites that use this discussion system.
Also Reddit has solved the problem you're mentioning (we've chosen the same solution, independently of each other).
- GGary Culliss?2013-03-24 03:04Z
I developed a similar technology for search results back in the late 1990's that considered how many times a result was selected by users relative to how many times a result was shown to users. We used similar methods of counting the links above a selected link as being shown to the user, as well pagination breaks, time and other metrics. You can then use ratios of the selections over the views for ranking, but the resulting number will be too volatile at low counts, so you need to build in some padding to account for that fact. Also, in search where results are more persistent, we took account of time by keeping track of the age of clicks and views, and expiring them after certain rolling periods of time. Anyhow, it's neat to see you come up with the idea for news. We applied it everywhere, including things like category listings and even to determine if something might contain inappropriate content.
- Uuser_145 (KajMagnus)2013-03-25 15:16Z
That's really cool :-) I googled your name and read about Direct Hit Technologies. I noticed you've studied engineering, law and finance — this seems almost a bit crazy :-) (in a good sense of course).
I found this article from 1999. I noticed you were using information about the user to decide which search results to show. Something interesting would be to use info about the user, to decide which comments to prioritize or highlight. For example, if the user is a rich middle aged man that votes on this or that party, then show highly upvoted comments that dissents with people from that group.
The intention would be to contribute to a more tolerant world, where people better understand others with different opinions.
(Perhaps this might even be doable, if one could gather information via Facebook or something. Hmm but there'd be lots of privacy issues of course. — Or people could contribute info about their own anonymous profiles, voluntarily, if they wanted to participate in "the-understand-others-better" project)
- MMatt?2013-03-24 05:49Z
Very smart solution. In fact, I hadn't recognized it was a problem until I read the title of your post... then it clicked immediately. Great idea & would definitely surface more interesting comments.
- Click to show more comments
- In reply toMatt:Uuser_145 (KajMagnus)2013-03-24 10:05Z
Oh, lucky choice of title then :-) I was wondering if it was too long
The problem discussed on this page might actually be a fairly recurring problem: People at Hacker News mentioned that search engines have a similar problem (that they've solved): they use link clicks (instead of upvotes) to estimate which search results are useful, but only the topmost search results tend to be clicked.
Edit: Also see Gary Culliss' comment to the left.
- Ffunflow?2013-03-24 07:32Z
Make reading comments a 2D navigation problem? Could be interesting
- Uuser_145 (KajMagnus)2013-03-24 09:46Z
It's a DAG graph actually I think :-) And a directed graph with cycles if one can edit one's comment and refer to what people said, later on.
I've been thinking about rendering a graph instead of a tree, but that feels... a little bit too crazy (and hard to implement)
- AAnonymous??2013-03-24 09:48Z
And as another commentor said, you also need to solve the problem with one malicious downvote.
- Uuser_145 (KajMagnus)2013-03-25 11:33Z
Measuring the screen time is a really interesting idea. Or rather, measuring how many people read a comment to the end, versus how many people only start reading it. But really hard to implement reliably? Except for mobile phones with a small viewport. Or if there was eye-tracking available :-)
- AAnonymous??2013-03-24 02:22Z
only 1 person sees your thread and upvotes = 100% so top comment for next person.
- Uuser_145 (KajMagnus)2013-03-25 10:59Z
Actually that's not how I'd implement it. Instead, one can use mathematics, to take into account how many people have read the comment: if only 1 person has read it, then we really don't know for sure if it is interesting, and we won't give it a score of 100%.
The mathematics stuff is called "using the lower bound of a binomial proportion confidence interval" — here's an article about it: How not to sort by average rating.
- Uuser_162 (Vijay)2013-03-24 02:37Z
This actually seems like a useful application of a multi-arm bandit algorithm. I definitely recommend John Myles White's O'Reilly book on MAB problems - it's brief and has some great discussion. A general concept is exploration vs. exploitation - using a view to explore a infrequently viewed comment, or exploiting a known good comment - or something in-between.
- Vvonskippy?2013-03-24 02:54Z
Upvotes/Downvotes are the bane of all discussion threads. What is this, Junior High? I can read and decide for myself if the comment is brilliant, valid, or just another jackass on the web. Just post the comments in the order they were made, and let the reader decide what to do with them. There's not a single website that uses upvotes/downvotes that hasn't turned into a huge circle jerk session of mutual back patting and popularity contests.
- Uuser_145 (KajMagnus)2013-03-25 11:52Z
What if there are 100 comments, or a forum topic with 100 pages filled with comments? Then people won't read them all and the interesting information is lost forever, somewhere in the middle...
Reddit, SlashDot, StackOverflow (!) + 100 StackExchange sites, and HackerNews are some sites that uses up/downvotes and work well. People do try to get more reputation points, and I think this tends to make them behave well and be more constructive and respectful towards each other.
- DDmitri??2013-03-24 03:26Z
I think it's a great idea. I've had similar thoughts before too, but I'm happy someone actually implemented (a nicely working, at first glance) solution.
- Uuser_164 (Matt)2013-03-24 04:59Z
I have thought of this much myself. I hope that commenting and voting continues to improve the way we digest discussions.
- Uuser_145 (KajMagnus)2013-03-24 11:38Z
If you'd like to think about it for many more days:
At Hacker News, whathappenedto posted links to long articles about how search engines handle a similar problem. — Perhaps many search engine techniques are actually relevant when it comes to sorting comments on a discussion site — if one views the comment sorting problem as a search for the most interesting comments.
- A4anonymous?2013-03-24 06:42Z
Doesn't this confuse long post with interesting post? Now a short, earth-shattering response will get no visibility vs long-winded replies...
- Uuser_145 (KajMagnus)2013-03-25 11:14Z
I'm not sure. Long posts do take longer to read, so fewer people tend to upvote them. I was thinking that, on the whole, the effect you're mentioning and the effect I'm mentioning cancel each other out.
However, my actual implementation might not work very well. Perhaps measuring which comment a visitor is reading, works well mainly on mobile phones (where only 1 comments is shown at a time).
I read an article by Paul Graham that discussed "good" and "bad" comments (the "Comments" section in the article), and he wrote that: "There is a strong correlation between comment quality and length", and "Whatever the cause, stupid comments tend to be short."
- PPaelidore??2013-03-24 06:51Z
This seems like an interesting solution, but what do you say to those who prefer a more "forum-based" format of reading? Would there be a more vertical version of this model?
- Uuser_145 (KajMagnus)2013-03-25 11:23Z
I think there ought to be — many people apparently like that format. But it'd probably take a while before I get/take the time to implement it. Perhaps someone else might want to do it, if/when Debiki becomes open source. (Under the AGPL license probably.)
- AAnonymous??2013-03-24 08:12Z
Too wide representation of comments. Super bad on mobile.
- S2sdoering?2013-03-24 10:00Z
Well, interesting idea. But I would propose that even people reading a comment, but doing nothing, would count as divisor for "upvoterate".
I know, would be difficult to achive and would have probably use some form of client-side-js to determine, if a comment had been visible in the browser. But seeing a comment and doing nothing, is a strong signal in itself. So you would get a picture more rooted in reality that way.
And yes, what you wrote is quite a problem for online-discussions, as the first-come first-win situation, does not push the best content to the top.
- Uuser_145 (KajMagnus)2013-03-25 11:39Z
Initially I was thinking about considering all visitors, like you suggested. Instead of considering only visitors who voted on something. But I'm afraid this would use up too much disk storage space (since the computer would have to store information on each and every visitor).
- BBastian?2013-03-24 14:29Z
well - i see two problems here:
if you have some collusion of people who upvote their comments in groups it would be very easy to dominate the comments section with new comments ... you have to factor in absolute values somehow because 1000ups/1000views is far better than 1up/1view
sorting this way will result in a very fluctuating comment section which probably isn't very usable - also most people dont take the time to equally consider all comments (not scrolling down e.g.) when giving upvotes ... maybe the solution should be to increase / decrease the font size of good / bad comments to make good stories more visible while still maintaining some order.
btw) going from left to right like on this side is equally unsuitable -- but the reddit-style comment system with threading and hiding bad comments is the best i know
- Uuser_145 (KajMagnus)2013-03-25 11:45Z
Re 1: One can use mathematics, to take into account how many people have read a comment. If only a few people have read it and upvoted it, then we really don't know for sure if it is interesting, and we won't give it a score of 100% interesting (1up/1view).
(This discussion system already works in that way actually :-))
Re 2: With problem 1 solved, I think this won't be that much of a problem.
Re btw: Fairly many people seem to prefer vertical layout rather than horizontal. I'm thinking about making this configurable per website.
- AAnonymous??2013-03-25 16:00Z
I think it's good, but I think that the replies to comments should be shown to the right of the original comment, rather than scrolling down to see them.
The reason for that is because it is still more convenient to scroll down than to scroll to the side, even with your fancy system in place. This is because of the page down and page up keys.
When a lot of low-content comments are posted, as is currently the case, it's important to have a great way to sift through them quickly. Threads, on the other hand, tend to need less space, and therefore can occupy the less searchable horizontal axis.
- HHopefully helpful??2013-03-26 02:20Z
The solution is simple - the weight of a vote should decay over time. It should affect up votes and down votes identically. This is more difficult to implement, but it is essentially what you're trying to approximate with your approach (number of reads being an approximation of the passage of time) and doesn't require trying to guess what's been read.
- Uuser_145 (KajMagnus)2013-03-26 03:58Z
That assumes people read all comments but they don't. Imagine a page with 1000 comments — most people read the topmost comments, but I'd guess that less than 1 in 1000 read all comments. Taking only elapsed time (or page views) into account, would unfairly favor the topmost comment with a factor of > 1000, and would not work well.
One needs to take both page views and position on screen into account.
It might, however, be a good idea to let a vote decay somewhat with time. Comments might get obsolete after some years (for example). So it might be a good idea to favor newly posted and popular comments, even if they're not quite as popular as a many years old comment. — I'd guess, however, that letting comments and articles be wikis, is good enough. (Then one can fix links that get broken, etcetera)