EVALUATION OF REPUTATION METRIC FOR THE B2C
e-COMMERCE REPUTATION SYSTEM
Anna Gutowska and Andrew Sloane
School of Computing and Information Technology, University of Wolverhampton
Wulfruna Street, Wolverhampton, U.K.
Keywords: Reputation System, B2C, e-Commerce, Simulation.
Abstract: This paper evaluates recently developed novel and comprehensive reputation metric designed for the
distributed multi-agent reputation system for the Business-to-Consumer (B2C) E-commerce applications.
To do that an agent-based simulation framework was implemented which models different types of
behaviours in the marketplace. The trustworthiness of different types of providers is investigated to
establish whether the simulation models behaviour of B2C E-commerce systems as they are expected to
behave in real life.
1 INTRODUCTION
The process of globalization creates new challenges
and opportunities for companies by offering an
access to new markets that were previously closed
due to cost, regulations etc. The adoption of the
Internet, in particular Internet-enabled B2C E-
business solutions, allows many Small and Medium
Enterprises (SMEs) to respond to these challenges
and opportunities by extending the geographic reach
of their operations. Very often however, Websites
created for sales purposes are simple in design and
functionality and therefore, do not arouse trust at
first glance. Furthermore, in contrast to “big brands”
which have already established their reputation in
the online marketplaces, SMEs are unknown to
many E-commerce customers.
In the E-commerce environment, which does not
require the physical presence of the participants,
there is a high level of “uncertainty” regarding the
reliability of the services, products, or providers.
Although many technologies exist to make the
transaction more secure, there is still the risk that the
unknown provider will not comply with the protocol
used. Thus, the decision of who to trust and with
whom to engage in a transaction becomes more
difficult and falls on the shoulders of the
individuals. In such an environment, reputation
systems come in place to assist consumers in
decision making. The basic idea is to let parties rate
each other to derive a trust or reputation rating. This
can assist in deciding whether or not to engage in a
transaction with this party. Reputation systems are
particularly useful in cases where the trustee is
unknown to the individual involved but well known
to others.
There are a number of existing consumer-to-
consumer (C2C) on-line reputation systems such as
eBay (2008) or Amazon (2008). However, unlike
C2C E-commerce marketplaces, most B2B sites do
not provide users with feedback information. There
are some centralized services/websites though,
which offer store ratings and reviews to their users,
such as BizRate (2008) or Resellerratings (2008).
All of them however, rely only on simple algorithms
calculating the average rating based on the given
feedback.
Nevertheless, much academic work on reputation
systems has been devoted to the C2C part of E-
commerce (Peer-to-Peer networks) which can be
found reviewed in (Sabater and Sierra, 2005; Josang
et al., 2007; Gutowska and Bechkoum, 2007a; Marti
and Garcia-Molina, 2006). Unlike the existing
centralized approaches (e.g. eBay, Amazon) which
are single-factor based, many authors proposed
distributed reputation systems which still tend to be
“one issue-centric” (Lin et al., 2005; Bamasak and
Zhang, 2005; Huynh et al., 2006; Fan et al., 2005)
(addressing only one of many problems existing in
the reputation systems (Josang et al., 2007;
Gutowska and Bechkoum, 2007a; Marti and Garcia-
489
Gutowska A. and Sloane A.
EVALUATION OF REPUTATION METRIC FOR THE B2C e-COMMERCE REPUTATION SYSTEM.
DOI: 10.5220/0001831104890498
In Proceedings of the Fifth International Conference on Web Information Systems and Technologies (WEBIST 2009), page
ISBN: 978-989-8111-81-4
Copyright
c
2009 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
Molina, 2006)). Even in studies attempting to
provide more complex reputation methods, for
example work on Histos/Sporas (Zacharia et al.,
2000), some issues are still not taken into
consideration, such as the transaction value, age of
rating, or credibility of referees.
Many of the problems addressed in C2C
reputation models also apply to the B2C E-
commerce environment. Not many authors however,
concentrate on the latter model of the marketplace.
The only work known to the authors addressing it is
(Cho et al.) and (Ekstrom and Bjornsson, 2002).
Nevertheless, to the best of the authors’ knowledge,
there are no studies whose main focus is to derive
reputation ratings in B2C E-commerce environment
taking into account the characteristics of the
providers.
This paper aims to evaluate a novel reputation
metric for computing reputation in a multi-agent
distributed B2C E-commerce system. To do that, the
agent-based simulation framework was
implemented. The strength of the metric is measured
by how well it reflects the agents (providers)
behaviour including resistance against different
hostile agents. Other important aspects of reputation
systems such as privacy of transaction data,
protection against collaboration attacks, and unfair
ratings are out of the scope of this work.
The proposed reputation metric evaluated in this
paper offers a comprehensive approach by including
age of rating, transaction value, credibility of
referees, and number of malicious incidents.
Furthermore, in addition to the information about
past behaviour it also incorporates other aspects
affecting online trust which are based on providers’
characteristics. Past behaviour, is not the only
information source affecting trust/reputation rating
of an online vendor. According to previous research
(Gutowska, 2007; Gutowska and Bechkoum,
2007b), there are many issues influencing online
trust-based decisions such as existence of trustmark
seals, payment intermediaries, privacy statements,
security/privacy strategies, purchase
protection/insurance, alternative dispute resolutions
as well as existence of first party information. The
extended approach evaluated in this paper yields a
promising improved distributed B2C reputation
mechanism. The problem of complexity of the
reputation metric is further described in (Gutowska
and Buckley, 2008a; Gutowska et al., to appear).
2 REPUTATION SYSTEM
FRAMEWORK
The simulated reputation metric in this study is
designed for the B2C E-commerce reputation system
in which two main roles are considered: buyer agent
i.e. agent representing a user and provider (web
service). A user agent collects for its user the
distributed reputation ratings about a web service
(provider). In return, a user provides the agent with
ratings about a transaction in order to build the
reputation database of the services. Agents create a
network where they exchange transaction ratings
about web services their users have dealt with, this
is called a buyers’ coalition in the paper. In this way
they are involved in a joint recommendation
process.
User agents and providers are engaged in a
transaction process e.g. buying-selling, where
money and products/services are involved.
To assess the reputation of a provider, first, a
user agent will use the information from the direct
interactions it has had with that party and second,
the ratings provided by other agents (indirect
ratings) from the buyers’ coalition which have dealt
with the provider.
The proposed reputation system is distributed
where each user agent will store their opinion about
transactions with other parties.
The assumption is that it is in users’ interest to
leave feedback after each transaction as that is the
only way the reputation system will work. The
participants are aware that if they want to calculate
the reputation rating about a particular provider the
only source of information will be their feedback
from the past and feedback received from other
users. Also, when entering the system as a new
member they are duty-bound to both. The users rate
transactions they were involved in and share this
information with others when requested.
3 REPUTATION METRIC
The reputation metric (Gutowska and Buckley,
2008a) evaluated in this paper is based on the
weighted average. The reputation value of provider
p is calculated as the arithmetic mean of the
compulsory reputation (Section 3.1) and the optional
reputation (Section 3.2). In addition the weight
wm(p) based on the number of malicious incidents is
applied.
WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies
490
If optional reputation metric is not calculated
then the reputation metric takes the value of
compulsory reputation metric multiplied by wm(p).
Further, the full rating scale of trust is [0; 1].
3.1 Compulsory Reputation
Compulsory reputation is based on the set of the
compulsory parameters and is defined as the
arithmetic mean of aggregated direct and indirect
ratings (see below). The rating scale for compulsory
reputation metric is [0; 1].
3.1.1 Compulsory Parameters
Transaction Ratings. In the proposed system the
quality factors constitute the explicit ratings left by
the user after the transaction and consist of three
components (as vector values): transaction outcome
i.e. if the product/service was received, fulfilling
provider’s signals (Rein, 2005) e.g. if the delivery
time, the product were as promised, as well as
customer service/support.
Raters’ Credibility (Implicit User Reputation).
Whilst choosing the group of users to require the
data from to calculate indirect reputation, it is
important to take their credibility as referees into
account. The reason for that is three-fold. Firstly, it
is often too costly or impossible to collect ratings
results from all interactions with the provider in
question (Josang et al., 2007). Secondly, to avoid the
inclusion of dishonest feedback into reputation
calculations from users demonstrating colluding
behaviour or leaving unfair ratings. Thirdly, to
choose the right subset of users with “similar
opinions”. Namely, different people have different
standards and they tend to trust the opinions of
people who have the same standards with
themselves (Zacharia et al., 2000). The solution
applied here is to extract users’ reputation
automatically and implicitly from their past
transaction rating data and use it to choose “n
best/most suitable raters”. The method presented
here is inspired by (Cho et al.) and uses raters’
ratings to estimate their underlying credibility. It is
based on the source credibility theory (Best et al.,
2003) which employs several schemes of
collaborative filtering methods (using similarities
between a target rater and the rest of the users). The
theory was shown to support rating mechanisms
both in the B2B (Ekstrom and Bjornsson, 2002) and
B2C (Cho et al.) E-commerce.
Source of Feedback. The reputation metric in this
study applies the weight ws(p) based on the rating
tendency concept proposed in (Cho et al.). It
decreases the rating from the rater who has a
tendency to rate higher than others, and vice versa.
)(1)(
)(uAu
gguws =
(1)
Where:
g
u
is the average transaction ratings from a rater u
g
A(u)
is the average ratings of the other users from
the subset of the “best/most suitable users”
(for the providers that the rater u rated).
Reputation Lifetime. In order to model the
dynamic nature of reputation, the weight associated
with the reputation lifetime wt is applied which
constitutes an exponential function of time. In this
way the more recent ratings are considered more
important and are valued higher comparing to the
older ones. Furthermore, as in (Zacharia et al.,
2000), the memory of the reputation system is
considered which disregards very old ratings.
)
(2)
( xt
x
wt
Δ
=
β
Where, t(x) is the time difference between the
current time (i.e. time of request) and the time when
the transaction x took place. β is used to scale t(x)
and β > 1. The time weight is applied to the
reputation metric in a recursive algorithm (Section
3.1.2).
Transaction Value. In counting reputation ratings
the value of the transactions is also taken into
account counteracting users who try to build a high
reputation by cooperating in many small transactions
and then cheat in a very large transaction. Also, the
transaction value range depends on the context to
which the reputation system will be applied i.e. the
maximum price of sold goods/services in the
marketplace. The weight associated with the
transaction value wv
x
is calculated using the
formula below:
)(
1
xv
x
wv
=
γ
and
0
1
= xwhere
vMax
x
γ
(3)
Where, v(x) is the value of transaction x and
vMax is the transaction range i.e. the maximum
value of the goods/services in the marketplace
(based on the context to which the reputation system
is applied). γ is used to scale v(x) and γ > 1.
EVALUATION OF REPUTATION METRIC FOR THE B2C e-COMMERCE REPUTATION SYSTEM
491
Number of Malicious Incidents. As in (Bamasak
and Zhang, 2005), in the proposed metric the
reputation value is reduced to the minimum when a
party reaches a certain threshold of malicious
incidents. Up to that threshold the appropriate
weight wm(p) is applied based on the exponential
function:
(4)
=
=<
0)(
)(0
pwmthenMmif
pwmthenMmif
m
α
where
0
1
= xwhere
M
x
α
(5)
Where, m is a number of malicious incidents of
provider p that occurred within the transactions
taken into calculation. M is the set threshold of the
number of malicious incidents above which the
reputation value is reduced to minimum. In the
equation above α is used to scale wm(p) and α > 1.
3.1.2 Computing Aggregated Ratings
The aggregated ratings are calculated with the
application of the recursive algorithm used on the
list of the transaction data records sorted according
to the time value.
The aggregated direct rating value is calculated
based on the data stored in the requesting agent a
database i.e. regarding its direct interactions:
[]
[
)(
)()()(
11)1(,
1,,
++
++=
xxxxa
xxxxaxa
wtwtwtAGRD
wtwtwtpURpAGRD
]
(6)
For the case where x=0 the aggregated direct
rating is equal to the updated rating for that
transaction. Where UR
a
,
j
(p) is the updated rating
value of transaction j with provider p calculated by
agent a
and x is the index of the last transaction on
the list (n-1).
The aggregated indirect rating values are
calculated in the same manner as above but are
based on the list of the transaction data from the
subset of the “n best/most suitable users”. In
addition, the weight ws is applied for each user
providing information.
3.1.3 Computing Updated Ratings
Updated reputation rating UR
a,x
(p) is calculated by
agent a for transaction x in which a was involved
with provider p
.
In general, each provider is reputed
by an agent after each transaction by providing a
transaction rating g. This is the average of two
components: fulfilling provider’s signals and
customer service, where both can take values [0; 1].
In addition, appropriate weight wv based on the
transaction value is applied.
3.2 Optional Reputation
In addition to the parameters presented above, a user
may choose to include some or all of the optional
parameters into calculations, which will influence
the rating value of a provider. They are: existence of
trustmark seals, existence of payment
intermediaries, existence of first party information,
existence of privacy statements, existence of
security/privacy strategies, existence of purchase
protection/insurance, and existence of alternative
dispute resolution and are further described in
(Gutowska and Bechkoum, 2007a; Gutowska,
2007).
The optional reputation is based on the set of
optional parameters (providers’ characteristics)
which take values [0; 1] and is presented by the
average of the above parameters which have been
chosen to be included into calculation. The rating
scale for optional reputation metric is [0; 1].
Optional reputation constitutes the initial
reputation for newcomers as at that point there is no
information of the past behaviour available.
4 SIMULATING B2C
e-COMMERCE REPUTATION
SYSTEM
The reputation system simulator used in this study
was developed in Java and it is based on a slightly
modified version of the RePast agent-based
simulation toolkit (Schlosser, 2004).
In the presented simulation the market is
populated by a number of agents that are divided
into buyers and providers. The simulation is based
on discrete time ticks. At each tick buyer agents are
supposed to initiate a transaction with a provider and
rate him afterwards. After the agents finished their
actions the data is collected and represented
graphically.
In the simulation the agents may enter or leave
the community with equal probability (see
Simulation parameters in Section 4.5.).
4.1 Modeling the Buyers
The buyers in the simulation framework differ in
types. The buyer agent type is a combination of its
trust disposition and its expectations.
WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies
492
Disposition to trust and the same risk attitude refer
to the fact that people have a baseline attitude when
they approach any trust situation. Some of them
have a higher baseline level of trust than others thus,
some individuals may find it easier/more difficult to
trust. The disposition to trust affects the decision of
either the buyer agent wants to engage in a
transaction with the provider or not (see the
acceptance function in Section 4.4.1.). Based on the
above there are different types of the buyer agents in
the simulation:
Risk Taking. This type of buyers is willing to take
risk easily which means they accept the high value
transactions even with the provider with low
reputation.
Very Cautious. This type of buyers is risk averse
and they are very careful with their decisions. They
accept the transactions only if the provider has high
reputation.
Conservative. Buyers representing this type come
between the two above extremes.
In the presented framework the buyer agents
have also different expectations towards the
outcome of the transaction which affects the way
they rate the transaction (see the rating function in
Section 4.4.2.). As in (Michalakopoulos and Fasli,
2005), there are three types of the buyers agents in
this study: optimists, realists, and pessimists.
Combining the two attributes discussed above
the following types of buyers agents were
implemented in the simulation framework: Risk
Taking Optimists, Risk Taking Realists, Risk
Taking Pessimists, Vary Cautious Optimists, Very
Cautious Realists, Very Cautious Pessimists,
Conservative Optimists, Conservative Realists, and
Conservative Pessimists.
4.2 Modeling the Providers
The effectiveness of a reputation system and its
metric depends on its resistance against malicious
behaviours. The success of non-honest agents is its
measurement for the quality of the metric (Schlosser
et al., 2005). Therefore there are different types of
providers implemented in the framework which are
called Trustworthy, Shady, Player, and Fly-By-
Night. They differ in their behaviour while
transacting (this is also correlated with their
characteristics). The characteristics of the interest
are the cheating probability (ChP) and the range of
the transaction outcomes they produce in terms of
customer service and fulfilling providers’ signals (in
other words the quality of services they provide).
The remaining attributes constitute the optional
parameters in the reputation metric and include:
existence of trustmark seals, payment
intermediaries, privacy statements, security/privacy
strategies, purchase protection/insurance, alternative
dispute resolutions as well as existence of first party
information. They have been chosen based on the
previously conducted survey discussed in
(Gutowska and Bechkoum, 2007a). The above
characteristics/optional parameters can take values
between 0 and 1 where 0 means no existence of the
attribute. In this way each type of the provider has
the optional reputation (OP) value based on the
above which constitutes the initial reputation value
for any new provider in the system. In the reputation
system there would be a devoted agent that would
gather the optional parameters information from the
providers’ Websites. The properties of different
providers are as follows:
Trustworthy. This type of the providers does not
cheat in the transactions (ChP=0) and provides high
service quality. All the parameters mentioned above
have high values (OP=0.92).
Shady. This agent does not have a particular pattern
in its behaviour (ChP=50). It provides false
statements on its Website which results in high
values of the optional parameters apart from
Trustmark Seals and Payment Intermediaries
(OP=0.63). The quality of the services it provides is
low.
Player. This type of a provider tries to build high
reputation by not cheating (ChP=0). When it
achieves its goal however, it starts behaving in a
malicious way (ChP:=100). When its reputation falls
down below the threshold then it starts being honest
again (ChP:=0). Player agent has got high values for
First Party Information, Privacy Statements and
Security Strategies (OP=0.43). When it does not
cheat the services provided are of a high quality.
Fly-By-Night. This agent’s goal is to cheat
(ChP=100). It provides false information about the
services it offers. The way of payment is direct to
the bank account (OP=0.51). The quality of the
services it provides is low.
4.3 The Simulation Cycle
The simulation framework is highly automated
where the handling of the agents, initiation of the
EVALUATION OF REPUTATION METRIC FOR THE B2C e-COMMERCE REPUTATION SYSTEM
493
transactions and storing the ratings are part of the
framework. The simulator repeatedly iterates a cycle
of events that would occur in the marketplace. The
steps of a transaction are as follows:
1. The simulation engine selects a buyer agent
who initiates a transaction with another provider
agent.
2. The buyer agent tests if the transaction is
acceptable i.e. he calculates the reputation of the
provider in question based on his previous direct
interactions as well as information from the
buyers community (the acceptance function is
described in Section 4.4.1.)
3. If the transaction takes place, the provider
agent determines the outcome of it and the buyer
agent rates it and stores the ratings. The ratings
depend on the buyer agent type and his
expectations and may not match exactly the real
outcome (the rating function is described in
Section 4.4.2.)
4.4 Modeling the Transaction
and Rating Processes
4.4.1 Transaction Acceptance Function
In the presented simulation the buyer agents have a
trust disposition which allows them to make
different decisions when it comes to engaging in a
transaction with a provider.
In this work the assumption is that no buyer
agent will transact voluntarily with a non-
trustworthy provider i.e. the provider with low
reputation. The other factor taken into consideration
while making the decision is the value of the
transaction. The acceptance function therefore, is a
correlation between the provider’s reputation and
the value of the transaction. The higher the value of
the transaction the higher the reputation should be
for the buyer to engage in this transaction. As
different people have different disposition to trust, in
the presented framework different types of buyer
agents have different acceptance functions. In this
way different types of agents accept the transaction
of a specific value at the different reputation level
Users’ willingness to trust however, can be
changed by experience (Shneiderman, 2000). In the
proposed framework all buyer agents representing a
specific type start with the same acceptance function
which is affected/changed later on by the outcome
of the transaction (experience) and in particular by
the providers’ malicious incidents. The calculation
of the acceptance threshold for a specific transaction
value with a specific provider is based on the
Lagrange Interpolation (Cheney and Kincaid, 1998).
4.4.2 The Rating Function
In the proposed framework each buyer agent rates
each transaction he has been involved in and collects
these ratings (see Transaction ratings Section 3.1.1)
in his database.
In a real marketplace, different people will rate a
transaction differently based on their experience and
their expectations towards the transaction outcome.
In the discussed simulation framework, three cases
are considered (as in (Michalakopoulos and Fasli,
2005)): optimists, realists, and pessimists. When it
comes to the transaction, optimists will be expecting
a very positive outcome, pessimists on the other
hand a rather bad outcome, and realists will come
somewhere between the two extremes. The
simulation framework addresses the above scenario
in a way that the optimist agent will hope for the
best outcome (in terms of customer service and
provider’s signals) he has had so far with the
provider in question, the pessimist agent will
anticipate the worst one, and the realist agent will
expect the average result based on his experience. If
the expected outcome (expOut) is higher than the
actual one (realOut), the buyer agent applies the
punishment value (p) to the transaction rating
(rating) which is a difference between the expected
and the real outcome value. If the expected outcome
value is equal or lower than the actual one, the
ratings reflect the outcome. The above rules are
presented below:
p:= expOut - realOut
if p>0 then
if p<= realOut then
rating:=realOut-Random(0, p)
else
rating:=realOut–Random(0,
realOut)
Apart from the transaction rating, the final
reputation value includes also the other component
which is Optional Reputation discussed in Section
3.2.
4.5 Simulation Parameters
There are several parameters which values can be
changed in the simulation framework depending on
the simulation needs. These are as follows: number
of starting agents (buyers and providers), number of
agents to add and remove in each step, add/remove
probability, the probability of initiating transaction
WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies
494
at each tick by a buyer agent, the amount of agents
of each type in the community, and number of
parallel simulations.
There are also some parameters which affect the
reputation metric itself. These are the maximum
number of malicious incidents, the transaction value
range, weight for the source of feedback and weight
for time factor. The maximum number of malicious
incidents is a threshold above which the provider’s
reputation is decreased to zero. The transaction
value range depends on the context to which the
reputation system will be applied i.e. the maximum
price of sold goods/services in the marketplace. The
above values determine the weights for the time
factor and malicious incidents of the reputation
metric which are based on the exponential function.
The weight for the source of feedback scales the
importance of the ratings coming from the indirect
interactions (buyers’ community) and the weight for
the time factor determines the impact of old ratings
vs. the more recent ones. This weight is also based
on the exponential function and is applied to the
reputation metric in the form of a recursive
algorithm (more details on the proposed reputation
metric in (Gutowska and Buckley, 2008b; Gutowska
and Buckley, 2008a)).
5 EVALUATION CRITERIA
The strength of the metric is measured by how truly
it reflects the agents (providers) behaviour and in
particular by its resistance against different hostile
agents. In the simulation the average requested
reputation, the market honesty, the acceptance rate,
the average number of transactions and the average
number of malicious incidents are calculated
separately for each type of the provider agents.
Average requested reputation is the mean
value of all reputation ratings of providers from a
specific type as if calculated/received by a buyer
when requesting reputation rating. This is based on
the rating information stored in the buyers’
databases.
Market honesty is the mean value of the actual
outcomes from the transactions produced by the
provider agents (not ratings). These are stored in
providers’ databases.
Acceptance rate is the proportion of
accepted/completed transactions with all initiated
transactions with providers of a specific type.
Average number of transactions is the average
number of transaction that a provider of a specific
type was involved in (accepted transactions).
Average number of malicious incidents is the
average number of malicious incidents for a specific
type of a provider.
The horizontal axis in the Figures below
represents the time. In Figures 1, 2 and 6 the vertical
axis corresponds to the computed reputation, in
Figures 3, 4 and 7 it represents the number of
transactions and in Figure 5 the acceptance rate.
6 SIMULATION RESULTS
As there is no work known to the authors that
introduces the reputation metric for the B2C E-
commerce reputation system, this paper presents
pioneering results and it does not compare the
efficiency of the evaluated reputation metric with
any other.
Table 1 presents the parameters and their values
used in the simulation.
Table 1: Parameters used in simulation.
The simulation results are depicted in Figures 1-
5. Market Honesty (Figure 1) and Average
Requested Reputation (Figure 2) show that the
reputation metric correctly reflects the behaviour of
different types of providers i.e. Trustworthy agents
keep their high reputation scores throughout the
experiment and the different types of malicious
agents have low reputation due to their transaction
history. It is noticeable that initially the reputation of
the malicious agents is a bit higher and it decreases
with time. This is caused by the fact that the initial
EVALUATION OF REPUTATION METRIC FOR THE B2C e-COMMERCE REPUTATION SYSTEM
495
reputation for new providers with no transaction
records is their optional reputation which in many
cases is based on the false information provided by
them on their Websites. When the transaction
information comes into the equation however, the
reputation algorithm appropriately deals with the
scenario and decreases the reputation value.
Figure 1: Market honesty.
Figure 2: Average Requested Reputation.
The slight difference in values between Market
Honesty and Average Requested Reputation reflects
the fact that different types of buyer agents rate the
transactions differently which does not always
match the real outcomes. The dissimilarity however,
is not significant which strongly suggests that the
reputation metric closely mirrors the behaviours in
the marketplace.
The results shown in Figures 3-5 indicate that
malicious agents are not involved in many
transactions (Figure 3) due to their low reputation.
The Acceptance Rate (Figure 5) decreases as the
buyer agents do not accept transactions with the
providers with low reputation. The Average Number
of Malicious Incidents (Figure 4) is kept stable
which is controlled by the Maximum Number of
Malicious Incidents simulation parameter. If the
parameter is set as M=1 then the reputation metric
will decrease the reputation of this provider to 0
which means he will not be accepted as a transaction
partner anymore and it will not get a chance to gain
profit by cheating. This scenario is illustrated in
Figure 6, ceteris paribus (i.e. while other parameters
stay unchanged).
Figure 3: Average Number of Transactions.
Figure 4: Average Number of malicious Incidents.
WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies
496
Figure 5: Acceptance Rate.
Figure 6: Average Number of Malicious Incidents, M=1,
ceteris paribus.
7 CONCLUSIONS
This paper presents and evaluates the novel,
comprehensive reputation metric designed for the
multi-agent distributed B2C E-commerce reputation
system. An agent-based simulation framework was
implemented that models the B2C E-commerce
marketplace. The results show that the proposed
reputation metric closely reflects different types of
behaviours in the marketplace. The method is
particularly resistant to malicious behaviour.
One of the assumptions of the proposed system
i.e., that there are no external parties included in the
framework can be easily amended in the future by
including the information coming from other
systems or reputation authorities. The other area
which could be looked at more closely is the
distribution of different buyer behaviours/types in
the real marketplace. The work on inclusion of those
in the proposed framework is underway.
REFERENCES
Am [01 July 2008].
Bam
vice, Hong Kong, 29 March -12
Bes
Building Marketing Strategy. McGraw-
Che cs
Cho
Systems with Applications, In Press, Corrected
eBa [01 July 2008].
Eks
ty.
Fan
s on
Gu
rmation Technology, University
Gu
lma de Mallorca, Spain, 29-31
Gu
ted Education, Dublin, 6-7 February,
Gu
azon (2008) [online].
<http://www.amazon.co.uk/>.
asak, O. & Zhang, A. N. (2005) A distributed
reputation management scheme for mobile agent-
based E-Commerce applications. In Proceedings:
IEEE International Conference on e-Technology, e-
Commerce and e-Ser
April,, pp. 270-275.
t, R., Corney, K. & Hawkins, D. (2003) Consumer
Behavior:
Hill/Irwin.
Bizrate (2008) [online]. <www.bizrate.com/>.
ney, W. & Kincaid, D. (1998) Numerical Mathemati
and Computing, International Thomson Publishing.
, J., Kwon, K. & Park, Y. Q-rater: A collaborative
reputation system based on source credibility theory.
Expert
Proof.
y (2008) [online].
<http://www.ebay.co.uk/>.
trom, M. & Bjornsson, H. (2002) A rating system for
AEC e-bidding that accounts for rater credibili
Proceedings of the CIB W65 Symposium, 753–766.
, M., Tan, Y. & Whinston, A. B. (2005) Evaluation
and design of online cooperative feedback mechanism
for reputation management. IEEE Transaction
Knowledge and Data Engineering, 17, 244-254.
towska, A. (2007) Research in Online Trust: Trust
Taxonomy as A Multi-Dimensional Model. School of
Computind and Info
of Wolverhampton.
towska, A. & Bechkoum, K. (2007a) A Distributed
Agent-based Reputation Framework Enhancing Trust
in e-Commerce. In Proceedings: IASTED
International Conference on Artificial Intelligence and
Soft Computing, Pa
August, pp. 92-101.
towska, A. & Bechkoum, K. (2007b) The Issue of
Online Trust and Its Impact on International
Curriculum Design. In Proceedings: The Third China-
Europe International Symposium on Software
Industry-Orien
pp. 134-140.
towska, A. & Buckley, K. (2008a) A Computational
Distributed Reputation Model for B2C E-commerce.
In Proceedings: IEEE/WIC/ACM International
EVALUATION OF REPUTATION METRIC FOR THE B2C e-COMMERCE REPUTATION SYSTEM
497
Gu
ational Conference on Distributed Computing
Gu ,
Special Issue on Trust and
Hu
nce on
Jos urvey of trust
Lin
l
Ma H. (2006) Taxonomy of trust:
Mi M.
Rei
l Hawaii
p. 26a.
Sabater, J. & Sierra, C. (2005) Review on Computational
A. dt (2004)
cumentation [online].
Sch n the
Zac Collaborative
reputation mechanisms for electronic marketplaces.
Decision Support Systems, 29, 371-388.
Conference on Intelligent Agent Technology
Workshops, 9 December, Sydney, Australia, pp. 72-76.
towska, A. & Buckley, K. (2008b) Computing
Reputation Metric in Multi-agent E-commerce
Reputation System. In Proceedings: IEEE
Intern
Systems Workshops, 20 June, Beijing, China, pp. 255-
260.
towska, A. Sloane, A. & Buckley, K. (to appear) On
Desideratum for B2C E-commerce Reputation
Systems. Journal of Computer Science and
Technology. The
Reputation Management in Future Computing Systems
and Applications.
ynh, T. D., Jennings, N. R. & Shadbolt, N. R. (2006)
Certified reputation: how an agent can trust a stranger.
In Proceedings: International Confere
Autonomous Agents and Multi-Agent Systems,
Hakodate, Japan, 8-12 May, pp. 1217-1224.
ang, A., Ismail, R. & Boyd, C. (2007) A s
and reputation systems for online servise provision.
Decision Support Systems, 43, 618-644.
, K. J., Lu, H., Yu, T. & Tai, C.-e. (2005) A reputation
and trust management broker framework for Web
applications. In Proceedings: IEEE Internationa
Conference on e-Technology, e-Commerce and e-
Service, Hong Kong, 29 March-1 April, pp. 262-269.
rti, S. & Garcia-Molina,
Categorizing P2P reputation systems. Computer
Networks, 50, 472-484.
chalakopoulos, & Fasli, M. (2005) On Different
Trust Attitudes and Their Effects on the Electronic
Marketplace. IEEE/WIC/ACM International
Conference on Intelligent Agent Technology, 102-108.
n, G. L. (2005) Reputation information systems: A
reference model. In Proceedings: Annua
International Conference on System Sciences, Big
Island, Hawaii, USA, 1-6 January, p
Resellerratings (2008) [online].
<
http://www.resellerratings.com/>.
Trust and Reputation Models. Artificial Intelligence
Review, 24, 33-60.
Schlosser, Technische Universitat Darmsta
Reputation System Simulator Do
<http://www.inferenzsysteme.informatik.tu-
darmstadt.de/users/schlosser/>.
losser, A., Voss, M. & Brückner, L. (2005) O
Simulation of Global Reputation Systems. Journal of
Artificial Societies and Social Simulation, 9, 4.
Shneiderman, B. (2000) Designing trust into online
experiences. Communication of the ACM, 43, 57-59.
haria, G., Moukas, A. & Maes, P. (2000)
WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies
498