hugo :: brainstorms in sociable media (mas.961 '03)
"only as an æsthetic phenomenon is
existence and the world justified"

- nietzsche


emotus ponens picture

::: reputation :::
To find out a girl's faults, praise her to her girlfriends. ~Benjamin Franklin


A game-theoretic analysis of the psychology of online reputation systems

Most online reputation systems, exemplified by ebay's seller and buyer feedback system, employ a signalling reputation paradigm. In these systems, the emphasis is on promoting trust, not on sanctioning bad behavior. Although users have the option of leaving negative feedback, empirical research (Resnick & Zeckhauser, 2002) suggests that negative feedback is often grossly underreported in these systems, which is the desired result for the ecommerce merchants hosting these reputation systems. It seems that the cause of this phenomenon is rooted in the publicity of the reputation feedback. Contrast signalling reputation systems with gossip-based reputation systems which lacks this publicity criteria. Because gossip about a person is decentralized over that person's social network, and the gossip data is often held in confidence and difficult to trace, negative reputation is allowed to flourish in the shadows. Machiavellianism would suggest that, given the opportunity to gossip without accountability, people would need very little incentive to spread negative gossip. Gossip-based reputation systems thus follow a sanctioning reputation paradigm, where negative feedback is sure to catch up with and can possibly be used to ostracize a person.

The publicity of reputation feedback in the ebay-style reputation system promotes positive feedback and resists negative feedback. 1) Users view giving positive feedback as social capital. Buyers giving positive feedback to sellers usually receive reciprocally positive feedback for themselves. Thus improving someone else's reputation will enhance one's own reputation. 2) Users also see positive feedback as a type of online social courtesy for a transaction, like saying, "Thank you" to the store clerk that bags your groceries. 3) Negative feedback is discouraged because the information is not anonymous and quite public. 4) Giving negative feedback could result in reciprocal negative feedback, a loss of opportunity to do future business with someone, and other forms of retribution. 5) Giving negative feedback is generally rude. 6) Sellers do not want to sell to buyers who are quick to give negative feedback.

Taking a game theoretic view, we can say that the ebay-style reputation system promotes cooperation. Even if a buyer or seller's experience is negative, it is to everyone's advantage to leave positive feedback and suppress negative feedback. Suppose a buyer purchases something and receives in late and damaged. The damage is done already. If he chooses to leave negative feedback, he risks receiving negative feedback at the most and will certainly lose the opportunity to gain positive feedback at the least. Thus, negative feedback leads to the same, or a lower reputation. There is no reward to the buyer for leaving negative feedback. If a buyer leaves positive feedback, she is likely assured of positive feedback. Thus, positive feedback leads to the same, but most likely, a greater reputation. There is no punishment for misrepresenting a negative experience with positive feedback. It is easy to see that the publicity of the ebay reputation system is the critical condition for making this system a signalling reputation system.

In contrast, gossip-based reputation systems are largely private, and thus, do not possess the accountability qualities of a system like ebay's. In a gossip-based reputation system, a person's reputation is not held publicly and centrally, but rather, is diffused throughout the person's social network. At every node in the diffusion of reputation information, the reputation information is likely to be corrupted or manipulated for competitive or political gain by the manipulator. "Gossip" is hard, but not impossible to trace. The psychology of gossip is also different from the psychology of leaving feedback in a central and public way. Saying something bad to someone's face is always harder than saying it behind their back because we do not feel as much accountability nor do we perceive any immediate threat of sanctions. In Information Warfare, Hess and Hagen suggest that gossip can be even more damaging when people mobilize into groups to collectively assault a person's reputation.

We can also take the game theoretic view of gossip-based reputation systems. Inherent in acommunity or social network are competitive pressures for resources or prestige. By gossiping positive reputation information about a person within the network, there is no expectation that if and when that positive reputation actually reaches that person, there will be any attribution to the originator of that feedback. Additionally, enhancing the reputation of another (unless both parties are within a coalition) will only cause one's own reputation to be relatively reduced. Thus, positive gossiping will not generate reciprocation because attribution is lost and will likely lower one's own reputation relatively speaking. By negative gossip, there is little risk that the originator or manipulator of the gossip will ever be known to the target of the attack. If the attack is successful, the originator or manipulator of the gossip can gain reputation relatively. Thus, negative gossip bears little risk to the gossiper and can lead to increased reputation for the gossiper (relatively speaking).

By comparing public reputation systems and gossip reputation systems used a game theoretic framework, we can see that the accountability of public reputation systems skews reputations to be more positive in general, whereas the lack of accountability and the competitive nature of social network gossip skews reputations to be more negative in general. Sanctioning in gossip reputation systems is somewhat explicit, whereas a "sanction" in a public reputation system is implicit and usually expressed as a user who is out-reputed by his competitors, failing to garner as much attention as them.

Hybridizing public and gossip reputation systems for eBay

Both public and gossip reputation systems have a place and role in ecommerce, and they are compatible with each other. Ebay's system, while promoting trust in the use of their markets, fails to reveal negative reputation information to potential consumers. The solution may be to embed a gossip reputation system within the existing public reputation system, which is still useful because it promotes trust in the community. Consider the following design.

User X wants to buy something from User Y. User X can read the public reputation infomation on User Y, but suspects there is more negative experience not being reported. User X broadcasts a request for information to all the other users who have dealt with User Y and have privately recorded negative feedback they chose not to publish. User Z's personal agent responds because User Z has had previous dealings with User X and trusts User X to not tattle-tale. On the basis of this trust, User Z's agent releases private feedback to User X, either signed, or anonymously. Included is User Z's own reputation. In this way, users can provide negative feedback to warn other potential buyers without fear of retribution. In fact, these users can still publicly come out as being positive on someone and enjoy the usual benefits of the public reputation system. User Z has motivation to release private reputation information to User X because there is a reciprocal agreement between the two users to share private information.

A positive testimonial/negative gossip scheme for reputation in online dating

Online dating can be a perilous forum. Participants search for potential matches and only have a potential mate's own self-description to go on. Often these self-descriptions are deceptive, and there are only so many assessment signals that can be read from such a profile. This is a forum where collaborative sharing of dating experiences with other members could be personally beneficial. However, the twist is that this forum is competitive by its nature. Bill will not want to share information with John on Mary because Bill and John are competitors. If Bill feels positive about Mary, he will want to conceal that from John. If Bill feels negative about Mary, he may actually want to pigeonhole John into a match with Mary to free up competition over other women. A central public reputation system will not suffice to provide warnings about negative reputation. If Mary has racked up negative feedback, she will simply leave the community or choose a new login name.

Just as with ebay, it seems that the best reputation system may be to maintain positive and negative information through separate systems. A positive testimonial system can be used to help males and females who have had positive dealings to cross-promote. They can enhance each other's positive reputations and these scores can be used to increase the prominence of a user in the results pages of potential suitors. However, a potential suitor cannot just go on positive reputation, since this only measures certain dimensions of reputation, such as social eptness. Also, it should be expected that such positive feedback will not be extremely informative. A male will want to leave positive testimonial for a female in hopes of reciprocation, but, unless he is completely uninterested in that person, he will probably leave fair vague praise.

Negative information must be handled differently, and a gossip system is preferable because there is no means of retribution. Negative gossip is treated as a commodity. It reveals dimensions of reputation not covered by positive reputation. In the following diagram, I illustrate the trade and barter of negative information.

The man labeled "81" is seeking gossip about the female. He receives only a fraction of the negative gossip on the female, proportionate to his "participation score" of 81. He has gained this score by previously recording feedback about his experiences with other women. Each time someone has accessed his feedback and judged it to be helpful and informative, his participation score has been rewarded. This is similiar to a measure of his reputation and expertise. Comparable systems exist in Amazon's book review system and's expert reviewers. Others are encouraged to accurately report the helpfulness of comments because if a person judges man "81" as not helpful, future gossip will be suppressed from this person. Men are encouraged to help out others because without being helpful, he himself will not receive help. Of course, a dyadic information bartering system would seem more fair, however, there may not be enough information reciprocity to motivate bartering between most men, e.g. it is unlikely that male X has info on female A and wants info on female B, and male Y has info on female B and wants info on female A. Thus, I opted for a community measure of participation.

The positive reputation system is motivated by cross-promotion and enables signaling, while the gossip reputation system is motivated by inside information as a commodity, and enables sanctioning/warning. By treating positive and negative reputation information separately in subsystems that best suit each, we can leverage their combined informational benefits.




H U G O . . L I U ...
interactive experience group

commonsense computing group
counter intelligence
hugo at media dot mit dot edu