> You see this anytime e.g. a high profile athlete "weighs in" on complicated geopolitical matters, when in reality their opinion on that matter should count next to nothing in most cases, unless in addition to being a great athlete they have also established a track record (reputation) of being expert or insightful in international affairs.
I apologize for multiple replies; I'm not stalking you. It's just an area I'm interested in and you're hitting on many ideas I've kicked around over the years.
I once got paid to write a white paper on a domain-based reputational system (long story!), based on just this comment. I think it requires either a formal taxonomy, where your hypothetical athlete might have a high reputation for sports and low one for international affairs, or a post-hoc cluster-based system identifies the semantic distance from one's areas of high reputation.
And reputation itself can be multi-dimensional. Behavior, like we've talked about elsewhere, is an important one. But there's also knowledge. Can the system model the difference betwwen a knowledgeable jerk (reputation 5/10) and a hapless but polite and constructive novice (reputation 5/10)?
So if your athlete posts about workouts, they may have a high knowledge reputation. And if they post about the design of stadiums, it's relatively closer to their area of high knowledge reputation than international affairs would be. And so on. And independently of their domain knowledge, they have a behavior reputation that follows them to new domains.
These are such great questions, and worth exploring IMO. My thinking is biased towards online communities (like old Usenet groups or mailing lists) and not towards giant free-for-alls like Twitter, so I think I have a lot of blind spots, but this is something I've thought about a lot too, so it's great to hear peoples' ideas, thank you.
> I think it requires either a formal taxonomy [...] or a post-hoc cluster-based system identifies the semantic distance
Yeah, I wonder if some existing subject-x-is-related-to-subject-y mapping could be used, at least as a hint to the system, e.g. all the cross-reference info from an encyclopedia. When communities become large enough, you might also be able to tease out a little bit of additional info from how many people in group X also participate in group Y maybe.
As an experiment, I'd also be curious to see how annoying it'd be to have your reputation not transfer across communities at all, but instead you build reputation via whatever that community defines as good behavior and having existing community members vouch for you (which, if you turn out to be a bad apple relatively soon after joining, then their endorsement ends up weakening their reputation to some degree a little too). There are some aspects to how it works in real life that are worth bringing over into this, I think.
> system model the difference betwwen a knowledgeable jerk (reputation 5/10) and a hapless but polite and constructive novice (reputation 5/10)?
I touched on this in a sibling comment somewhere though I've long lost track of the threads, but I think the platform would want to rely on human input to some degree - part of being a good community member is doing things like periodically reviewing a batch of posts that got flagged by other users. In one community, there might be an 'anything goes' mentality, where another may set stricter standards around what's considered normal discussion, and so I think it'd be hard for a machine to differentiate but relatively easy for an established community member (again though, it always has at least a micro impact on your reputation, so how you carry out your reviewing duties can also increase or decrease your reputation in that community).
Odds are too that, if you occasionally have to put on the hat of evaluating others' behavior, it might help you pause next time you're tempted to fly off the handle and post a rant.
Anyway, the focus of the technology would be less about automatically policing behavior and more about making it easier for communities to to call out good and bad behavior without much effort, and then having that adjust a person's reputation - often very tiny adjustments that over time accumulate to establish a good reputation.
> independently of their domain knowledge, they have a behavior reputation that follows them to new domains
I apologize for multiple replies; I'm not stalking you. It's just an area I'm interested in and you're hitting on many ideas I've kicked around over the years.
I once got paid to write a white paper on a domain-based reputational system (long story!), based on just this comment. I think it requires either a formal taxonomy, where your hypothetical athlete might have a high reputation for sports and low one for international affairs, or a post-hoc cluster-based system identifies the semantic distance from one's areas of high reputation.
And reputation itself can be multi-dimensional. Behavior, like we've talked about elsewhere, is an important one. But there's also knowledge. Can the system model the difference betwwen a knowledgeable jerk (reputation 5/10) and a hapless but polite and constructive novice (reputation 5/10)?
So if your athlete posts about workouts, they may have a high knowledge reputation. And if they post about the design of stadiums, it's relatively closer to their area of high knowledge reputation than international affairs would be. And so on. And independently of their domain knowledge, they have a behavior reputation that follows them to new domains.