Applied ArtiÐcial Intelligence, 14:881È907
Copyright 2000 Taylor & FrancisÓ
0883-9514
/
00 $12.00
1
.00
TRUST MANAGEMENT
THROUGH REPUTATION
MECHANI SMS
GIORGOS ZACHARIA and PATTIE M AES
MIT Media Laboratory,
Cambridge MA, U.S.A.
The me mbers of electronic communities are often unrelated to e ach other; they may have
never met and have no information on each otherÏs reputation. This kind of information is
vital in electronic commerce interactions, where the potential counterpartÏs reputation can
be a signiÐcant factor in the negotiat ion strategy. Two complementary reputation
mechanisms are investigated which rely on collaborative rating and personalized evaluation
of the various ratings assigned to each user. While these reputation mechanisms are
developed in the context of electronic commerce, it is believed that they may have
applicability in other types of electroni c communities such as chatrooms, newsgroups,
mailing lists, etc.
““ Although an application designerÏs Ðrst instinct is to reduce a noble
human being to a mere account num ber for the computerÏs convenience, at
the root of that account number is alw ays a human identity.ÏÏ (Khare &
R ifkin, 1997)
Online com munities bring together people geographically and s ociologi-
cally unrelated to each other . Online communities have traditionally been
created in the context of discussion groups, in the form of newsgroups,
mailing lists, or chatrooms. Online communities are usually either goal or
interest-oriented. But other than that, there is rarely any other kind of bond
or real-life relationship among the m embers of communities before the
members meet each other online. The lack of information about the back-
ground, character, and especially the reliability of the members of these com-
munities causes a lot of suspicion and mistrust among their members.
When a newcomer joins a chatroom, a news group, or a mailing list,
he
/
she does not know ho w seriously he
/
she should take each participant
until he
/
she has formed an opinion about the active members of the group.
Likewise, the old members of th e group do not know how seriously they
should take a newcomer until he
/
she establishes him
/
herself in the group. If
Add ress correspondence to Giorgos Zacharia, MIT Media Laboratory, 20 Ames Street, E15-305,
Cam bridge, MA, 02139, USA. E-mail: lysi@media.mit.edu
881
882 G. Zacharia and P. Maes
the group has a lot of traffic, the noise-to-signal ratio becomes too high, and
the process of Ðltering out the interesting messages becomes increasingly
difficult for a newcomer or an occasional reader of the group. If users did
have an indication for the reputation of the author of each message, they
could prioritize the messages according to their predicted quality.
Similar problems are encountered in other kinds of online communities.
The recent development of online auction sites, and other forms of electronic
marketplaces has created a new kind of online community, where people
meet each other to bargain and transact goods. Online marketplaces like
Amazon Auctions (Amazon), Kasbah (Chavez & Maes, 1996), MarketMaker
(Wang, 1999), eBay (eBay), and OnSale Exchange (OnSale) introduce two
major issues of trust:
buyers have no physical access to the product of interest while
d
Potential
they are bidding or negotiating. Therefore, sellers can easily misrepresent
the condition or the quality of their products.
sellers or buyers may decide not to abide by the agreement
d
Additionally,
reached at the electronic marketplace, asking later to renegotiate the
price, or even refuse to commit the transaction. Even worse, they may
receive the product and refuse to send the money for it, or the other way
around.
Although thes e problems of trust are als o encountered in real-world
experiences, the problem is more difficult in online communities, because
one has very few cues about other people by which to evaluate them. Many
of the signals that we use in real life ar e absent in online environments, and
thus alternative methods of adjudicating reputation are needed.
One way of solving the above-mentioned problems in the system would
be to incorporate a reputation brokeri ng mechanism, so that each user can
customize his
/
her pricing strategies according to the risk implied by the
reputation values of his
/
her potential counterparts.
Reputation is usually deÐned as the amount of trust inspired by a partic-
ular person in a speciÐc setting or domain of interest (Marsh, 1994). In
““Trust in a Cryptographic EconomyÏÏ (Reagle, 1996), reputation is regarded
as asset creation and it is evaluated according to its expected econ omic
returns.
Reputation is conceived as a multidimensional value. An individual may
enjoy a very high reputation for his
/
her expertise in one domain, while
having a low r eputation in another. For example, a Uni x guru will probably
have a high rank regarding Linux questions , while he may not enjoy as high
a reputation for questions regarding MicrosoftÏs operating systems. These
individual reputation standings are developed through social interactions
among a loosely connected group that shares the same interest. Also, each
R eputation M echanisms 883
user has his
/
her personal and subjective criteria for what makes a user repu-
table. For example, in the context of a discussion group, some users prefer
polite mainstream postings, while others engage in Ñame wars. Through this
interaction, the users of online communities establish subjective opinions of
each other.
Methods have been developed through which one can automate the
social mechanisms of reputation for electronic communities. An early
version of these reputation mechanisms has been implemented in Kasbah
(Chavez & Maes, 1996). Kasbah is an ongoing research project to help
realize a fundamental transformation in the way people transact goods-
from requiring constant monitoring and e†ort, to a system where software
agents do much of the bidding and negotiating on a userÏs behalf. A user
wanting to buy or sell a good creates an agent, gives it some strategic direc-
tion, and sends it o† into the marketplace. Kasbah agents proactively seek
out potential buyers or sellers and negotiate with them on their creatorÏs
behalf. Each agentÏs goal is to make the ““best dealÏÏ possible, subject to a set
of user-speciÐed constraints, such as a desired price, a highest (or lowest)
acceptable price, and a date to complete the transaction (Chavez & Maes,
1996). In Kasbah, the reputation values of the individuals trying to buy
/
sell
books
/
CDs are major parameters of the behavior of t he buying, selling, or
Ðnding agents of the system.
The second section of this paper describes the related work in the
domain of rating systems and reputation mechanisms. The third section out-
lines the requirements for a successful reputation mechanism for online com-
munities. The fourth section describes problems speciÐ c to electronic
marketplaces and online discussion forums. The Ðfth and sixth sections
describe two reputation mechanisms that have been designed and evaluated.
The seventh section evaluates the mechanisms using simulations and user
data from eBay and Amazon auctions. The last section is the conclusion of
the paper and the outline of future work.
RELATED WORK
The related work on reputation system s can be divided into two major
categories : noncomputational reputation systems like the Better Business
Bureau Online (BBB) and computational ones. The Better Business Bureau
Online is a centralized repository of consumer and business alerts. They
mainly provide information on how well businesses handle disputes with
their clients. They also keep records of the complaints about local or online
companies and even publish consumer warn ings agains t some of them. They
do not provide any kind of numerical ratings for business or consumer trust-
worthiness.
884 G. Zacharia and P. Maes
The computational methods cover a broad domain of applications, from
rating of newsgroup postings and w ebpages, to rating people and their
expertise in speciÐc areas. This section focuses on the related computational
methods and a comparison of their major features (Table 1).
One way of building a reputation mechanism involves having a central
agency which keeps recor ds of the recent activities of the users of the system,
very much li ke the scoring systems of credit history agencies. The credit
history agencies use cus tomized evaluation mechanisms provided by the
software of FairIsaac (FairIsaac) in order to ass ess the risk involved in
giving a loan to an end consumer. The ratings are collected from the pre-
vious lenders of the consumers, and consumers are allowed to dispute those
ratings if they feel they have been treated unfairly. The resolution of a rating
dispute is a responsibility of the end consumer and the party that rated the
particular consumer.
However useful a centralized approach may be, it requires a lot of over-
head on behalf of the service providers of the online community. Further-
more, the centralized solutions ignore possible personal affinities, biases, and
standards that vary across various users.
Other p roposed approaches like Yenta (Foner, 1997), Weaving a Web of
Trust (Khare & Rifkin, 1997), and the Platform for Internet Content Selec-
tion (PICS), such as the Recreational Software Advisory Council (RSAC),
are more distributed. However, they require the us ers to rate themselves and
to have either a central agency or other trusted users verify their trustwor-
thiness. One major problem with these systems is that no us er would ever
label him
/
herself as an untrustworthy person. Thus, all new members would
TABLE 1 Comparison of Online Reputation Systems. In the ““Pairwise RatingÏÏ
Co lumn It is Indicated Whether the Ratings are Bi-Directional or
One -Directional, and Who Submits Ratings. In the ““Personalized EvaluationÏÏ
Co lumn It is Indicated Whether the Ratings are Evaluated in a Subjective Way,
Base d on Who Makes the Query
Pe rsonalized Te xtual
S ystem Pair-wise rating Evaluation co mments
F ireÑy Rating of recommendations Yes Ye s
G roupLens Rating of articles Yes No
W eb of Trust Transitive ratings Yes No
e Bay Bu yers and sellers rate each No Ye s
o ther
Am azon Bu yers and sellers rate each No Ye s
o ther
O nSale Bu yers rate sellers No Ye s
Cre dit history Le nders rate customers No Ye s
P ICS Self-rating No No
E lo & Glicko Resu lt of game No No
B izrate Con sumers rate businesses No Ye s
R eputation M echanisms 885
need veriÐcation of trustworthiness by other trustworthy users of the system.
In consequence, a user w ould evaluate his
/
her counterpartÏs reputation by
looking at the numerical value of his
/
her reputati on as well as the trustwor-
thiness of his
/
her recommenders.
Yenta and Weaving a Web of Trust introduce computational methods
for creating personal recommendation systems, the former for people and
the latter for webpages. Weaving a Web of Trust relies on the existence of a
connected path betw een two user s, while Yenta clusters people with
common interests according to recommendations of users who know each
other and can verify the assertions they make about themselves. Both
systems require prior existence of social relationships among their users,
while in online marketplaces, deals are brokered among people who may
have never met each other.
Collaborative Ðltering is a technique for detecting patterns among the
opinions of di†erent users, which can then be used to make recommen-
dations to people, based on opinions of others who have show n similar
taste. This technique bas ically automates ““word of mouthÏÏ to produce an
advanced and personalized marketing scheme. Examples of collaborative Ðl-
tering systems are HOMR, FireÑy (Shardanand & Maes, 1995), and Group-
Lens (Resnick et al., 1994). GroupLens is a collaborati ve Ðltering solution
for rating th e contents of Usenet articles and presenting them to the user in
a personalized manner. In this system, users are clustered according to the
ratings they give to the same articles. These ratings are used for determining
the average ratings of arti cles for that cluster.
The Elo (Elo, 1978) and the Glicko (Glickman, 1999) systems are compu-
tational methods used to evaluate the playerÏs relative strengths in pairwise
games. After each game, the competency score of each player is updated
based on the result and previous scor es of the two users. The basic principle
behind ratings in pairwise games is that the ratings indicate which player is
most likely to win a particular game. The probability that the stronger
player will win the game is positively related to the di†erence in the abilities
of the two users. In general, the winner of a game earns more points for
his
/
her rating, while the defeated player loses points from his rating. The
changes in the ratings of the two users depend on their rating di†erence
before the game takes place. If the w inner is the player w ho had a higher
score before the game, the change in the ratings of the two users is nega-
tively related to their rating di †erence before the game. If, however, the
winner of the game is the player who had a lower score before the game
took place, the changes in the sores of the two players are positively related
to their rating di†erence before the game.
BizRate (BizRate) is an online shopping guide that provides ratings for
the largest 500 companies trading online. The ratings ar e collected in two
di†erent ways. If BizRate has an agreement with an online company, t he
