Fixing Hacker News with a reputation system
There is a certain phenomenon that seems to happen in almost every online community of user-generated content. A community is created: the initial users define the values of this new community. After a while the community experiences growth in numbers. As a result of that growth, users that joined before it feel like its no longer the same community with the same values. The latest widely discussed example seems to be Hacker News.
Paul Graham responds that the reasons are mostly: a shift in values, increase of anonymity and the fact that its easier to vote than to contribute content
It's a genuine problem and has been growing gradually worse for a while. I think the cause is simply growth. When a good community grows, it becomes worse in two ways: (a) more recent arrivals don't have as much of whatever quality distinguished the original members, and (b) the large size of the group makes people behave worse, because there is more anonymity in a larger group.
I've spent many hours over the past several years trying to understand and mitigate such problems. I've come up with a bunch of tweaks that worked, and I have hopes I'll be able to come up with more.
The idea I'm currently investigating, in case anyone is curious, is that votes rather than comments may be the easiest place to attack this problem. Although snarky comments themselves are the most obvious symptom, I suspect that voting is on average dumber than commenting, because it requires so much less work. So I'm going to try to see if it's possible to identify people who consistently upvote nasty comments and if so count their votes less.
As online communities grow, the values of the group shift. The majority now may or may not hold the same values as the majority before. The question is, how to preserve the old values of the group with minimum side-effects?
As it happens, my master's thesis was an attempt to fix exactly this problem and implement an improved voting system tailored specifically for communities with user-submitted content. I won't provide a link to the thesis as its not written in English, but I'll try to summarize the gist of it.
The voting system used in most communities today (democratic voting) is the one most susceptible to value shift when significant growth occurs. Its no surprise: democratic systems are designed to measure what the majority values. When significant growth occurs, the majority changes and therefore what they value also changes.
In contrast, previous moderator/editor based systems offer a strict filter on content based on the more static values of the current set of editors. However, it has the downside of being limited to what the editors are able to review and publish.
I propose a hybrid feedback-loop based system. In this system people have variable voting influence and editor-like individuals are given as a "reference point" or exemplary users with maximum voting influence. The system attempts to find out what they value and recognize it in others.
The system is largely based on the beta reputation system, which is used to measure trust in online e-commerce communities.
Here is a short description of the system:
- Voting influence is not the same for all users: its not 1 (+1 or -1) for everyone but in the range 0-1.
- When a user votes for a content item, they also vote for the creator (or submitter) of the content.
- The voting influence of a user is calculated using the positive and negative votes that they've received for their submissions.
- Exemplary users always have a static maximum influence.
Suppose we have a content item submitted by the user . Now a voter comes to vote for it and clicks on the +1 button.
The voter has their own submissions for which they has received a total amount of positive vote and a total amount of negative vote . As a result, their voting influence is modified: its not +1 but calculated according to the formula:
where is the lower bound of Wilson score confidence interval. While a simple average such as:
might work when the number of positive and negative votes is large enough, its not good enough when the number of votes is low. The Wilson score confidence interval gives us a better, flexible balance between desired certainty in the result and the result itself.
This vote in turn is received by the content item . Because its a positive vote, the amount of positive vote is changed for this content item
and as a result, it has a new rating
but the positive points of the creator of the content item are also changed:
and as a result the voting influence of submitter is also changed:
or in other words, the submitter "earned" a bigger influence in the voting system by submitting a well-rated content item.
This means that new members have no voting influence. As they submit content and receive votes their influence may rise if the existing users with high influence in the system consider their content to be good.
This is where the reference users come in. Their influence is fixed to always be 1
Because of this, influence propagates through the system from them to other users who submit content which is deemed high-quality by the reference users. Those users in turn also change influence by voting for others and so forth.
Its also possible to scale down votes as they age. The two possible strategies are to scale all and values daily, for all content items and all users by multiplying them with a certain aging factor
or to simply keep all positive and negative votes and in the database and recalculate and according to the age of the votes , for example:
One of the convenient aspects of this system is that its easy to test-drive. It doesn't require more user action than simple democratic voting. It only requires an administrator to specify some reference users at the start which seed and then propagate influence throughout the system.
I tested this system on a forum dataset (details available on request) and found that the system achieves around 50% reduction of difference from a moderator only system compared to scores of a democratic system, even when the direct voting of reference users is turned off for content items and only the indirect (to other users) influence is counted.
What does a 50% reduction in the difference mean? Let the score of a content item be measured in 3 systems: democratic , reference-users-only and hybrid with direct influence of reference users to content items being turned off. By sorting the items according to those scores we can calculate their ranks in the 3 systems: , and respectively. The value of the rank is in the range to , where is total number of content items. The absolute difference between the democratic ranking and the reference ranking is:
while the absolute difference between the hybrid ranking and the reference ranking is:
and it turns out that on average,
The important downside of these results is that the people using the system were not aware that points are calculated in a different way. The original votes were given by people who knew that the system is democratic and acted accordingly. It remains to be seen what the results would be if people are aware that their voting influence depends on the way others vote for their submitted content.
I pondered starting a website similar to hacker news based on this voting and scoring scheme, however starting a whole new news website is about much more than just scoring algorithms (it requires reputation in the online comminty, popularity and most importantly time, none of which I presently have in sufficient amounts or know how to achieve). But hopefully, pg and the rest of the hacker news team might find this scheme useful enough to somehow incorporate it into the existing scoring system.