A Bayesian Reputation System for Virtual
Organizations
Jochen Haller
SAP Research
Vincenz-Priessnitz-Str. 1,
76131 Karlsruhe, Germany
jochen.haller@sap.com
Abstract.
Virtual Organizations (VOs) are an emerging business model in to-
day’s Internet economy. Increased specialization and focusing on an organiza-
tion’s core competencies requires such novel models to address business oppor-
tunities. In a VO, a set of sovereign, geographically dispersed organizations tem-
porarily pool their resources to jointly address a business opportunity. The deci-
sion making process determining which potential partners are invited to join the
VO is crucial with respect to entire VO’s success. The possibility of a VO partner
performing badly during the VO’s operational phase or announcing bankruptcy
endangers the investment taken in integrating their processes and infrastructure
for the purpose of the VO. A reputation system can provide additional decision
support besides the a priori knowledge from quotations and bidding to avoid
events such as VO partner replacement by helping to choose reliable partners in
the first place. To achieve this, reputation, an objective trust measure, is optimally
aggregated from multiple independent trust sources that inherently characterize
an organization’s reliability. To allow for the desired predictions of an organiza-
tion’s future performance, a stochastic modeling approach is chosen. The paper
will present a taxonomy of TIs for VO environments, a stochastic model to main-
tain and aggregate trust sources, so called Trust Indicators, and the inclusion of
other subjective measures such as feedback.
1 Introduction
In today’s business world, commercial relationships become increasingly flexible and
are formed on demand whenever a business opportunity emerges. To cater for these
changing demands, business relationships integrate Information and Communication
Technologies (ICT) to (partially) automate certain decision processes, e.g. the swift dis-
covery and selection of business partners. Virtual Organizations (VOs) are a prominent
example of such emerging business models. A VO is defined as a temporary coalition
of otherwise independent organizations or individuals, collaborating to achieve a com-
mon business goal, one party alone could not master. A VO follows a phased life cycle,
first, a VO manager or system integrator observes a business opportunity and discovers
potential VO partners (identification phase). Second, the VO manager negotiates with
the potential partners, until the required set of VO members is selected, bringing the
right set of specialized expertise in the VO. During this phase, the ICT infrastructure
Dagstuhl Seminar Proceedings 06461
Negotiation and Market Engineering
http://drops.dagstuhl.de/opus/volltexte/2007/999
in each partner’s domain is set up and configured for collaboration in the VO and also
contracts are established (formation phase). Third, the VO executes, each VO member
runs his business processes contributing to the overall VO goal - to exploit the business
opportunity - until it is reached (operational phase). In case of abnormalities or excep-
tions like misperforming VO members, adaptations can be made by the VO manager
(evolution phase), but the VO stays within the operational phase. Finally, fourth, af-
ter having achieved the VO goal, common goods, VO results etc. are dispersed among
the VO members according to the contractual agreements. Final processes, for instance
billing, are performed (dissolution phase) [1, 2].
VO like structures can already be observed in different business domains. While
on the one hand established domains, such as automotive engineering still retain rather
static business relationships, a car manufacturer for instance tends to stay with well
established and confirmedly reliable car part suppliers, recently founded domains on
the other hand exhibit more agile buyer-supplier relationships. An example for the lat-
ter is the high-tech industry, such as chip manufacture, where chip prices change on
a daily basis. When a buyer selects a supplier, this essentially means taking a trusting
decision. If the buyer has no prior knowledge of a particular (set of) supplier(s), there
is no absolute certainty about the supplier’s future reliable behavior. In that case, an
online reputation system can help to minimize the risk of selecting a bad supplier who
frequently delivers late or not at all. In contrast to hard security measures that apply
during regular operations between a buyer and supplier such as confidential communi-
cation channels or access control measures, reputation based decision support belongs
to the class of soft security measures [3]. Soft security aims at complementing hard se-
curity rather than replacing it. In the context of this publication, we focus on centralized
reputation models that cater best for reputation and trust requirements in VOs [4]. The
reputation system is assumed to be owned or hosted by an explicitly trusted third party
(TTP). Following up on the above statement about the buyer’s expectation of a supplier,
we define trust as the subjective probability by which the buyer expects a supplier to
behave reliably [5]. While trust is a subjective probability, different buyers have differ-
ent trust perceptions or expectations of the same buyer, reputation strives at providing
an objective trust measure. It is defined as the business context specific aggregation of
(subjective) trust values from multiple independent sources to support a supplier’s deci-
sion making process with respect to an intended collaboration with prospective buyers
[6].
In the following, this short paper will outline a snapshot of work currently conducted
on a reputation system model that caters for the following, not necessarily only VO
specific, requirements:
1. integrate a stochastic trust management model that takes the specifics of business
relationships into account
2. in particular, tie trust to observable parameters (in the following termed as Trust
Indicators, short TI), inherently characterizing the abilities of a business partner
3. deal with uncertain, incomplete business partner information in dynamic VO envi-
ronments
4. take long running VOs into account that require all involved parties to query for
reputation, not only a buyer investigating about suppliers
2
5. integrate direct feedback to increase the quality of future reputation responses
Section 2 will introduce the reputation model, that integrates a stochastic trust man-
agement approach where trust is tied to a taxonomy of so-called trust indicators char-
acterizing business partners. A feedback mechanism will also be outlined. Section 3
concludes.
2 Trust and Reputation Model
Many existing trust management approaches, especially in Peer-2-Peer (P2P) environ-
ments, root trust solely in feedback given by peers after having conducted a transaction
[7, 8]. Feedback is a highly subjective, relative input that may vary between peers who
participated in the same transaction. In contrast to these approaches, we believe that
rooting trust in absolute, observable properties - the TIs - inherently characterizing an
organization’s reliability is a sounder approach for trust among VO members. Feedback
is still an important input that improves the quality and accuracy of the reputation value
aggregating the TI rooted trust values.
2.1 Taxonomy of TIs
This subsection provides an excerpt of an assembled classification/taxonomy of TIs
and their individual modeling. The full taxonomy with more extensive examples will
be published in a full paper soon. A TI models one aspect of trust in an organization,
participating in VO. Since VOs can emerge in different business domains, e.g. high-tech
or engineering industries, having a large set of members with specialized expertise, TIs
can have heterogeneous origins and meaning. Therefore, it makes sense to classify TIs
that share a similar origin and meaning. This approach has the benefit that the rele-
vance of a TI class for a particular business domain and even individual VO members
requesting reputation values can be easily determined.
Operational TIFinancial TI Organizational TI External TI Third Party TI
Trust IndicatorTrust Indicator
Fig. 1. TI Taxonomy
Figure 1 shows the top-level TI classes of a taxonomy currently encompassing 146
different TIs. TIs are classified into:
–
Financial TIs modeling the financial trust aspects of an organization, an example is
the cash flow quote indicator [9].
–
Organizational TIs modeling the organizational reliability and stability of an orga-
nization, e.g. with the employee fluctuation indicator [10].
– Operational TIs model an organization’s operational reliability, for instance with
the delivery delay indicator [11, 12].
3
– External TIs cater for trust relevant aspects external to an organization such as with
the country bond spread index indicator that aggregates country risk [13].
– Third Party TIs allow for interfacing with other specialized, third party reputation
or expert systems providing trust relevant indicators in form of recommendation.
It already becomes obvious from the TI top-level class descriptions that the taxon-
omy exercise is an interdisciplinary effort drawing, among others, from the fields of
risk management, operations research and Key Performance Measurement. A related
approach was conducted by Tan [14] who assembled a Trust Matrix entailing trust as-
pects for Electronic Commerce but who remained with his work on a higher and abstract
level.
After the top-down description of the high level TI classes, we follow with the
detailed modeling of individual TIs. Re-iterating, a TI is an observable property char-
acterizing a trust aspect of an organization, therefore with an impact on its reputation.
Taking the operational TI ”delivery delay” as an example, the time difference passing
between an agreed upon delivery date by buyer and supplier and the actual delivery
date is observed. Most suppliers aim at minimizing delay, hence it can be expected that
suppliers more often deliver slightly late with the number of suppliers decreasing with
increased delivery delay.The goal of TI modeling is finally to predict future VO member
behavior based on previously collected (TI) data. The fact that the availability of cor-
rectly observed data can not be guaranteed in productive systems motivates a TI model
based on probability distributions. The described behavior of the ”delivery delay” TI
for instance suggests a model based on the Exponential distribution [6]. To cater for
missing or incorrect data, observed data does not directly determine the TI distribution.
Instead, a Bayes update, by applying the following Bayes theorem equation is used:
P (θ|X) =
P (X|θ)P (θ)
P (X)
=
P (X|θ)P (θ)
P
S
P (X|θ)P (θ)
Observed data X contributes to the prior or empirical distribution P (X). The dis-
tribution assumption, in case of delivery delay that it is exponentially distributed, deter-
mines the likelihood distribution P (X|θ) estimating the true parameter θ. Evaluating
the equation leads to the posterior distribution P (θ|X) with the best fit of the distribu-
tion parameterized by the observed data to the distribution assumption. The posterior
distribution is further employed for aggregating towards a reputation value. Assuming
discretionary density functions, allowing for more efficient numerical evaluation, the
right side of the equation is evaluated. The normalizing denominator’s sum index S
iterates over the equidistant intervals or states of the grid.
Figure 2 visualizes the described example distributions graphically, based on data
sampled from a Enterprise Resource Planning system.
Besides these mathematical properties, the TI model also entails the following at-
tributes:
Name N.
Every TI is uniquely identified by a name N.
Domain D.
A TI can be based on observations of a continuous or discrete variable x.
The possible values of x are the domain of the TI.
4
Fig. 2. TI Delivery Delay
Update time period ∆t
upd
.
Trust information is likely to arrive in different intervals.
The attribute ∆t
upd
defines a fixed time grid telling the reputation system, how
often to update a TI.
Observation time period ∆t
obs
.
The time period ∆t
obs
defines a maximal time win-
dow to look into the past. Beyond that, observations are regarded to carry no more
significance.
Time weighting function ω. Among n observations x
i
at times t
i
, i ∈ {1, ..., n} within
the time window, old ones are less likely to carry information about future values
than newer ones. Each TI incorporates a monotonically increasing weighting func-
tion ω(t) > 0 that can implement forgetting of older observations and put emphasis
on newer ones.
Trust preference mapping π.
In order to judge the level of trustworthiness displayed
by a TI, we define an ordinal scale 1 to p
max
, where 1 represents the lowest and
p
max
the highest level of trust indicated by the TI. To compare TIs, the scale is
the same for all TIs. p
max
= const. π defines a function π : S → {1, ..., p
max
}
mapping the states S to the different levels of trust indicated by them. This mapping
enables an expert to incorporate his knowledge on the particular TI domain.
It has to be noted that this information rich TI model can express complex proper-
ties of business transactions and relationships. Other reputation systems in related work
already attempted stochastic models for business transactions or returned feedback, but
these attempts could only model binary events or transactions, that is if a transaction
ended positive or negative. A prominent example, the Beta Reputation System in [15]
uses the Beta distribution that is parametrized by the amount of positive and negative
previous outcomes of a particular transaction. Real business transactions involving de-
livery of chip components such as heatsinks, do not behave that atomically and may be
long-running, requiring a more sophisticated model of transaction indicators and their
attributes.
5