Gaming the System

Hey @Narrative Team and fellow Narrators!

    I was staring at the accrued NRVE reward bucket the other day and had a bit of an unsettling thought.  Has the @Narrative team thought about safeguards as to how 'popular' content is ranked?  Perhaps this all ties into the reputation algorithm, which makes me feel better.  The immediate concern is of course someone 'viewing their own post a ton of times'...hopefully unique views are observed in some way, along with figuring out how to avoid people 'paying for views via click farms' like they do for youtube and other mediums.  One of the major issues of course is that 'public people without accounts' will not have reputations.  Our model is also slightly different than other platforms as well, since anyone can plainly see what kind of rewards are at stake for the current payout cycle.

What do the team and others think about this?

Original Post

Hi @Banter

I understand - and I think all the system can do to mitigate this is to assign less weight to views from people who are not Narrators.  

Similarly to the reputation boost that comes with having done KYC, a weight boost should differentiate those who have signed up to Narrative and therefore have a reputation, and those who have not, and so could be part of a click farm operation.

So we're compounding measures against three levels of risk:

1) the non Narrator who could be part of a very large click farm and has no rep score

2) the Narrator who has a rep but may not be a unique user because no KYC 

3) the intra-Narrative click farm: Narrator accounts that can be KYC verified but that still may exist solely for the aim of gaming the system

For Narrators with or without KYC a further differentiation could be made for those who post content that is deemed above a certain minimum quality threshold or within an acceptable quality range (mainly to weed out people who might be spamming with random posts just to try to meet this criteria).  This would be useful because although their reputation already takes post quality into account, it should carry extra weight for the specific consideration of whether a Narrator is a good actor or potentially part of an intra-Narrative click farm.

Hey @Malkazoid, I agree with your general assessment / proposed solution.  It will be interesting to see what the team finally settles on.  I feel like any post that does really well (at least in the early days), will be 'viral' outside of Narrators with accounts, which won't necessarily benefit the poster all that much depending on how the algorithm is structured.  If I know that 100 outside views ~ 1 view from a kyc narrator with good reputation, I might spend my time focusing on 'marketing' to my fellow narrators rather than external marketing.  This mindset will ultimately hurt the spread of Narrative going forward if the incentive to 'spread' the word isn't as great as hyping up your content within the existing community.  In a perfect world, I guess we would do both, but there is only so much time

Drixx Madison posted:

Is it set in stone that posts will be "view" based rather than "Like" based? That distinction may make a difference.

Good question!  I have no idea and am actually only assuming views will be part of the computation.  If it were my design, I would be looking at both views and likes, with likes weighted heavier than views.  How much heavier can only be properly determined through trialing the algorithm but I guess I would start at 3 or 4 times heavier.

I believe that ultimately ties back to your point about Narrators (people with reputation, having more weight).  I could be wrong, but I don't believe the general public will have any ability to interact with a post (like, comment, etc), they would need an account for that.  Anyhow....I just wanted to toss the question out there, I'm sure the team has thought about it some, and just wanted to gain some insight into how content would ultimately get 'ranked'.

Good point, Banter, I think this sort of ties in with another post I just made that people might create a ton of junk content for their niches and/or try to use that junk content to boost their visibility and clout on Narrative. Right now three of the top 10 niches by number of posts are "Audi," "Nike," and "Air Jordan," three brands that I wouldn't think would attract much beyond marketing material, affiliate links, and empty content from enthusiasts talking about how much they absolutely adore their favorite athletic shoes. Maybe we could work it into the reputation system that they can be dinged if there are signs that they're trying to artificially increase their clout through things like fake traffic?

Drixx Madison posted:

Is it set in stone that posts will be "view" based rather than "Like" based? That distinction may make a difference.

I have always been under the assumption that it is "like" based, not views. Given that your participation in this platform gets you activity points, It always made sense to me that rewards would be accounted for based on votes for the content. That is why i haven't commented on this thread yet because I was confused why we were talking about views. 

It will be interesting to see how the metrics are laid out. But I suspect that since the platform is so much about transparency, that the actual voting up (or maybe even down) is a far more transparent approach, than view count which individuals can't truly verify.

Heather Aston posted:

Good point, Banter, I think this sort of ties in with another post I just made that people might create a ton of junk content for their niches and/or try to use that junk content to boost their visibility and clout on Narrative. Right now three of the top 10 niches by number of posts are "Audi," "Nike," and "Air Jordan," three brands that I wouldn't think would attract much beyond marketing material, affiliate links, and empty content from enthusiasts talking about how much they absolutely adore their favorite athletic shoes. Maybe we could work it into the reputation system that they can be dinged if there are signs that they're trying to artificially increase their clout through things like fake traffic?

That is why i suspect it is based on the same vote up or vote down system that our niche approval has.

 

Banter posted:

Hey @Malkazoid, I agree with your general assessment / proposed solution.  It will be interesting to see what the team finally settles on.  I feel like any post that does really well (at least in the early days), will be 'viral' outside of Narrators with accounts, which won't necessarily benefit the poster all that much depending on how the algorithm is structured.  If I know that 100 outside views ~ 1 view from a kyc narrator with good reputation, I might spend my time focusing on 'marketing' to my fellow narrators rather than external marketing.  This mindset will ultimately hurt the spread of Narrative going forward if the incentive to 'spread' the word isn't as great as hyping up your content within the existing community.  In a perfect world, I guess we would do both, but there is only so much time

I hadn't considered that - you're right, and the algorithm will have to strike a balance between the opposing imperatives of wanting to demote external views that are inherently more likely to be an attempt to game, and wanting narrators to market their content externally to the network.

At first there will be a natural motivation for all of us to market externally though, since the low number of internal users at the outset will mean there just won't be that many internal views to court.  I suggest the @Narrative Network Team includes a variable that adjusts for this effect, and reduces its value over time to model as best as possible the increasing pressure to market internally as the internal market grows.

 

Malkazoid posted:
Banter posted:

Hey @Malkazoid, I agree with your general assessment / proposed solution.  It will be interesting to see what the team finally settles on.  I feel like any post that does really well (at least in the early days), will be 'viral' outside of Narrators with accounts, which won't necessarily benefit the poster all that much depending on how the algorithm is structured.  If I know that 100 outside views ~ 1 view from a kyc narrator with good reputation, I might spend my time focusing on 'marketing' to my fellow narrators rather than external marketing.  This mindset will ultimately hurt the spread of Narrative going forward if the incentive to 'spread' the word isn't as great as hyping up your content within the existing community.  In a perfect world, I guess we would do both, but there is only so much time

I hadn't considered that - you're right, and the algorithm will have to strike a balance between the opposing imperatives of wanting to demote external views that are inherently more likely to be an attempt to game, and wanting narrators to market their content externally to the network.

At first there will be a natural motivation for all of us to market externally though, since the low number of internal users at the outset will mean there just won't be that many internal views to court.  I suggest the @Narrative Network Team includes a variable that adjusts for this effect, and reduces its value over time to model as best as possible the increasing pressure to market internally as the internal market grows.

 

I don't think the external views, needs to be factored into your rewards though. I mean, sharing it outside of the platform is a given, because if you already have an interested audience in your particular voice, you will by human nature be interested in enticing them to join Narrative, as they are mostly already in a favorable position to follow and vote up your content once they join. I think this is plenty of motivation.

Emily Barnett posted:
Heather Aston posted:

Good point, Banter, I think this sort of ties in with another post I just made that people might create a ton of junk content for their niches and/or try to use that junk content to boost their visibility and clout on Narrative. Right now three of the top 10 niches by number of posts are "Audi," "Nike," and "Air Jordan," three brands that I wouldn't think would attract much beyond marketing material, affiliate links, and empty content from enthusiasts talking about how much they absolutely adore their favorite athletic shoes. Maybe we could work it into the reputation system that they can be dinged if there are signs that they're trying to artificially increase their clout through things like fake traffic?

That is why i suspect it is based on the same vote up or vote down system that our niche approval has.

 

@Emily Barnett - if you are suggesting that the outcome of content approved to a niche could cause the moderator who approved it to receive a small ding if it falls below a certain level of quality, I agree that could be an interesting route to explore.  The process would have to be quite forgiving, with the ding being small, and the threshold being set to only cause demerit for very low value posts... since the idea is to only produce a penalty for behaviour that is deliberately manipulative.

As will probably always be the case, it will be important to use a sliding scale of demerit proportional to just how beyond the threshold the post is.  But the system could also monitor a moderator's content scores, and warn them if there seems to be a pattern of questionable moderation - if the problem does not improve within a week of the warning, the demerit could jump into a higher range. 

 

Niche owner involvement

To involve a human element into this, without involving the Tribunal, we could consider that the niche owner could be a good person to involve in this process. 

The warning could first be sent to the niche owner, giving them some choices as to how to proceed.  They could be given a link to a digest of the moderator's approval choices so their human eyes can provide a human analysis of the gravity of the situation.  They could then choose between various degrees of increased demerit, ranging from a mild increase in the penalty, all the way to the suspension of the moderator for a period of time (which would carry a significant reputation ding as well).

Once the niche owner has chosen the appropriate remedy, the warning could then be sent to the moderator communicating the problem, and the outcome if their performance does not increase within a set time period (perhaps a week).

Over time, if the system flags the same moderator more than once with this sort of problem within a given time window (perhaps six months months), then the options given to the niche owner in terms of demerit could increase in severity, including the ability to terminate the moderator's position.

I've been concerned that niche owners should have that option if they get paired with a problematic moderator.  They need some recourse.  This could be one approach to providing such recourse.

This brings about another consideration - a period report to the owner of moderator performance could be a great way to keep owners involved in their niches.  Even if all their moderators are performing above system-gaming levels, knowing which ones are admitting the lowest quality content, and which ones are getting poor or great feedback from the community on their moderation performance, could help owners in their interactions with their mods.  They could be given the ability to influence slightly the distribution of moderator rewards to their mods, based on performance.  For instance let's say 70% of the mod rewards could be guaranteed to be distributed automatically via whatever calculation Narrative determines, but the remaining 30% could be performance based, and skewed in favour of the best mods in a distribution pattern the niche owner gets to determine.  And perhaps the niche owner could be given the ability to determine the percentage of guaranteed rewards, within a range of 50% to 90%?  A 50% setting would then create the strongest performance-based setting, and 90% would create an environment where the performance bonus is small?

I realise these considerations are drifting away from those of prevention of system gaming, and more towards improving the system overall... but still interesting!

 

Emily Barnett posted:

I don't think the external views, needs to be factored into your rewards though. I mean, sharing it outside of the platform is a given, because if you already have an interested audience in your particular voice, you will by human nature be interested in enticing them to join Narrative, as they are mostly already in a favorable position to follow and vote up your content once they join. I think this is plenty of motivation.

External views still remain part of evaluating how many non-Narrative eye balls are being brought to the platform via a piece of content.  There is a value inherent in that. 

I completely agree there is also a natural motivation to share beyond the platform, but how effectively you do so should be rewarded to some degree.  I trust that over time the algorithms will home in on how to do so optimally.

Likes are way more important than views for judging content value, but the best algorithms incorporate variables for as many factors as possible, and views alone are a factor, as well as likes.  Their value should be included in the formula.

A distinct section of the algorithm should also consider number of likes per view - since that is perhaps the easiest reflection of pure quality to calculate.  If your piece has 1000 views, and 250 likes, it is a fair call to determine to some degree that that post is higher quality than one that gets 2000 views, and 250 likes.

Malkazoid posted:

@Emily Barnett - if you are suggesting that the outcome of content approved to a niche could cause the moderator who approved it to receive a small ding if it falls below a certain level of quality, I agree that could be an interesting route to explore.  The process would have to be quite forgiving, with the ding being small, and the threshold being set to only cause demerit for very low value posts... since the idea is to only produce a penalty for behaviour that is deliberately manipulative

Yes exactly @Malkazoid, I think just like all activity impacts up or down, on your reputation, Moderators should have this too. Perhaps if there is a real obscure reason why the moderator felt it was relatable to the niche and allowed it to be published, they can provide a comment immediately explaining their thought process, just as we sometimes see on the niche suggestions.

It is also good to remember that the community can vote a moderator out of the position if they are sloughing off, and not actually moderating.

Malkazoid posted:
Emily Barnett posted:

I don't think the external views, needs to be factored into your rewards though. I mean, sharing it outside of the platform is a given, because if you already have an interested audience in your particular voice, you will by human nature be interested in enticing them to join Narrative, as they are mostly already in a favorable position to follow and vote up your content once they join. I think this is plenty of motivation.

External views still remain part of evaluating how many non-Narrative eye balls are being brought to the platform via a piece of content.  There is a value inherent in that. 

I completely agree there is also a natural motivation to share beyond the platform, but how effectively you do so should be rewarded to some degree.  I trust that over time the algorithms will home in on how to do so optimally.

Likes are way more important than views for judging content value, but the best algorithms incorporate variables for as many factors as possible, and views alone are a factor, as well as likes.  Their value should be included in the formula.

A distinct section of the algorithm should also consider number of likes per view - since that is perhaps the easiest reflection of pure quality to calculate.  If your piece has 1000 views, and 250 likes, it is a fair call to determine to some degree that that post is higher quality than one that gets 2000 views, and 250 likes.

I look at Medium as a good example. You can share your content in order to spread the word of the platform and your content, but those views don't count for anything in determining monthly payouts (at least not to my knowledge) The monthly payouts are pro rata based on how many member's claps, your content accumulates..

How do you propose quality views  without engagement? I mean if views are calculated in, it is incredibly easy to game the system, one just has to start up a pay for view app just like Youtube and Instgram have with pay per likes. $5 bucks gets you 1000 page views. 

If any reward was to be calculated from views, then i would only feel positive if the metrics were about views that resulted in conversion being rewarded. Did the person sign up up to Narrative? Otherwise views are more about advertising benefits rather than strengthening our digital ecosystem.

I truly hope monthly rewards are only calculated based on actual membership activity. It encourages people to join vs just consume content. IMO

Emily Barnett posted:

I look at Medium as a good example. You can share your content in order to spread the word of the platform and your content, but those views don't count for anything in determining monthly payouts (at least not to my knowledge) The monthly payouts are pro rata based on how many member's claps, your content accumulates..

I'm not familiar with what happens behind the scenes for them either.

How do you propose quality views in without engagement? I mean if views are calculated in, it is incredibly easy to game the system, one just has to start up a pay for view app just like Youtube and Instgram have with pay per likes. $5 bucks gets you 1000 page views. 

Easy: external views would have to be valued way below what is economically achievable with third party view providers.  I'd imagine an article on Narrative which gains 1000 external views might be paying out 95% based on its internal Narrative metrics of Narrator views, likes and comments, with only a small percentage of its value derived from external views.  This should make it virtually impossible for it to be economically viable to buy external views and profit with NRVE.

If any reward was to be calculated from views, then i would only feel positive if the metrics were about views that resulted in conversion being rewarded. Did the person sign up. Otherwise views are more about advertising benefits rather than strengthening our digital ecosystem.

I truly hope monthly rewards are only calculated based on actual membership activity. It encourages people to join vs just consume content. IMO

Agreed of course.  As a separate consideration to the value of views alone, conversion should certainly have a much higher value.

I think @Malkazoid and @Emily Barnett and others are bringing up good points.  I'm looking forward to seeing what the @Narrative team has to say about it.  As I'm the one that started this thread, one of my original concerns was 'paying for views', and how that will be mitigated.  On the flip side, I'd hate to see such a heavily skewed formula to value outside eyeballs at 5% vs 95% for internal views. To use my favorite example, let's say I create some amazing, revolutionary macaroni art content, and I work hard to market it externally, and it goes viral, like big time.  The vast majority of those eyeballs coming to view the article are going to be non-narrators that won't be able to interact ( like the post / comment), unless they sign up.  However, once advertising is in place, those views are going to be invaluable to the platform.  As such, I feel like I should get compensated accordingly. 

    I feel like one of the potential problems with valuing 'internal narrator' interaction so highly is nepotism.  Let's say I'm someone with a ton of followers on the platform.  And lets same the unspoken rule is that if you view / like / interact with my content, I'll do the same for you, a little quid pro quo.  I feel like this would tend to promote the most popular people's content regardless of the quality or more wide-spread public appeal.

    Since Medium has been around for a while now, it really would be interesting to know how they value external views for content payout.  If it were truly just a matter of 'internal claps', it seems like it could be gamed by the scenario I laid out above.

    Speaking of places like youtube and what not, clearly advertisers aren't shying away from videos that are getting 'click farmed'.  I'm sure they are factoring in the real views vs fake views when paying for ads these days.  So at the end of the day it comes down to what @Malkazoid mentioned, there has to be some factor that we multiply incoming views by to get at their 'real value' like an advertiser would for the platform.  I just don't think it is as low as 5%

Banter posted:

I think @Malkazoid and @Emily Barnett and others are bringing up good points.  I'm looking forward to seeing what the @Narrative team has to say about it.  As I'm the one that started this thread, one of my original concerns was 'paying for views', and how that will be mitigated.  On the flip side, I'd hate to see such a heavily skewed formula to value outside eyeballs at 5% vs 95% for internal views.

Hey @Banter... as I mentioned, opposing external to internal views, I'd start with a 3:1 or 4:1 valuation ratio.  So that's either 33.33% vs 66.66%, or 25% vs 75%.

Then I'd run a platform simulator which I imagine the dev team is including in its suite of tools, and see what kind of values the economy develops from these settings... and adjust from there.

Notice where I mentioned 5% vs 95%, I was looking at external views vs all internal Narrative valuations (views, likes, comments, and who knows what else).

In a 4:1 external to internal view valuation ratio, this would mean:

- 20% of the post value would be derived from views both internal and external, with 5% of the value coming from external views, and 15% from internal views

- 80% of the post value would be derived from the sum of values computed from likes and comments, and any other value marker generated by Narrative users)

Advertisers won't value internal or external views any differently.  So we should consider that separately to the question of value of views to ward against gaming the system.

For advertisers, what we need to decide is how we assign value to views, vs likes, vs comments, etc.  

Once we've arrived at the value for advertisers, we can factor that into a larger calculation that takes into account imperatives to prevent gaming.

So to return to a scenario where 20% of valuation comes from views, and 80% from likes, comments and other Narrative user valuation factors, that 20/80 split seems like a potential initial setting in our simulations, considering views will probably be the lowest value of all interactions considered.  This would be appropriate for an established Narrative community of significant size.  The variable factor we discussed earlier that would change with community size could then adjust backwards in time to balance the system for its initial size at launch...

@Banter To me the outside views negates the reputation we are all building, quite a bit. If we are judging the content as a portion, for eyeballs on the page then you are admitting the quality is only a percentage (As @Malkazoid suggests maybe 75%). Something being viral does not guarantee that the content is of higher quality. In fact quite often it is the opposite. I think this could take us in a lesser direction, that far out weights the problem of nepotism. Remember, only last year, eating tide laundry pods went viral. Not quality content! I would be very sad if that kind of content was scooping up the rewards because  25 percent of it's worth came from fresh eyeballs on it. Not the kind first impression I was hoping for the platform. Of course we have no real controls over what content is made and goes viral. But the outside viewing rewards, is a double edge sword in my opinion. 

I see the nepotism issue being only a small problem for a small time. As the platform grows, communities will divide and bloom in small little gardens of content makers, mainly around topics that people are interested in.

I also agree with @Malkazoid that the advertising is a separate issue than how we are rewarded. Case in point. Your Macaroni Art post may just take off like viral colours, but a last the platform has yet to attract either Catelli or Kraft as advertisers yet, so no ads are even on this viral post (because advertisers have determined the  niche is not relative to their brand). So what are we really gaining by the outside views? If the post doesn't inspire people to join narrative, what are we actually rewarding the viral macaroni art content for?

I say, outside views need to be measured in real conversion to membership, or else viral doesn't have much clout in my opinion. 

Also gaming the system for that hypothetical 25% percent of outside viewership, for NRVE rewards, will be rampant in a blink of the eye. BOTS

 

Random thoughts: How far would the reputation system affect someone who can churn out content and also get a lot of his buddies to join Narrative and upvote his content? Would it somehow undo the reputation system if someone could create a ton of bot accounts for the pure purpose of upvoting their content? Should there be some kind of machine learning system in place that can look for signs that a niche might be riddled with spam so that niche can be made less visible until the owner and/or moderator can get it sorted out? Perhaps the system could add in total time spent and engagement level with a post alongside votes and raw number of views?

Heidi Hecht posted:

Random thoughts: How far would the reputation system affect someone who can churn out content and also get a lot of his buddies to join Narrative and upvote his content? Would it somehow undo the reputation system if someone could create a ton of bot accounts for the pure purpose of upvoting their content? Should there be some kind of machine learning system in place that can look for signs that a niche might be riddled with spam so that niche can be made less visible until the owner and/or moderator can get it sorted out? Perhaps the system could add in total time spent and engagement level with a post alongside votes and raw number of views?

I believe all of what you suggested here is exactly the directions the Narrative team has in mind. They have said from the get go that bot voting will quickly loose reputation. I believe the activity of being present and active on the platform in human activity is what gets you activity rewards. I suspect that is part of the reason why we are doing kyc, that you get an added reputation boost for proving you are a legit person, not a bot. We don't know enough about moderation yet. Like if tagging to a niche is instantaneously posted and the mod job is to remove inappropriate content, or the other way around...that tagged content forms a que until moderators can put eyes on it and then release to the niche.

 

The next month should be interesting as things get revealed to us.

Emily Barnett posted:
Heidi Hecht posted:

Random thoughts: How far would the reputation system affect someone who can churn out content and also get a lot of his buddies to join Narrative and upvote his content? Would it somehow undo the reputation system if someone could create a ton of bot accounts for the pure purpose of upvoting their content? Should there be some kind of machine learning system in place that can look for signs that a niche might be riddled with spam so that niche can be made less visible until the owner and/or moderator can get it sorted out? Perhaps the system could add in total time spent and engagement level with a post alongside votes and raw number of views?

I believe all of what you suggested here is exactly the directions the Narrative team has in mind. They have said from the get go that bot voting will quickly loose reputation. I believe the activity of being present and active on the platform in human activity is what gets you activity rewards. I suspect that is part of the reason why we are doing kyc, that you get an added reputation boost for proving you are a legit person, not a bot. We don't know enough about moderation yet. Like if tagging to a niche is instantaneously posted and the mod job is to remove inappropriate content, or the other way around...that tagged content forms a que until moderators can put eyes on it and then release to the niche.

 

The next month should be interesting as things get revealed to us.

True. One thing I'd like to see is the owner and moderator coordinate on what the "goals" of a niche are. Like, if the owner spots something that might only be tangentially related to the niche, he or she can talk it over with the moderator to have it removed. That way, we have less of niches getting hijacked even by accident because people are tagging their posts to the wrong niches. Or should that be limited so the owners and moderators aren't removing stuff just because they disagree with an opinion?

@Heidi Hecht and others. Good discussion going on here.  Just a few thoughts - it will be hard for someone to get their buddies to join only to upvote one person's content. New folks need to earn reputation before their votes count more than someone who has already earned a higher reputation. Reputation impacts  the value of the vote.  And, we do plan on incorporating machine learning later. 

Add Reply

×
×
×
×