REFINING THE TRUSTWORTHINESS ASSESSMENT OF

SUPPLIERS THROUGH EXTRACTION OF STEREOTYPES

Joana Urbano, Ana Paula Rocha and Eugénio Oliveira

LIACC – Laboratory for Artificial Intelligence and Computer Science, Faculdade de Engenharia da Universidade do Porto

DEI, Rua Dr. Roberto Frias, 4200-465, Porto, Portugal

Keywords: Situation-aware Trust, Dynamics of Trust, Multi-agent Systems.

Abstract: Trust management is nowadays considered a promising enabler technology to extend the automation of the

supply chain to the search, evaluation and selection of suppliers located world-wide. Current agent-based

Computational Trust and Reputation (CTR) systems concern the representation, dissemination and

aggregation of trust evidences for trustworthiness assessment, and some recent proposals are moving

towards situation-aware solutions that allow the estimation of trust when the information about a given

supplier is scarce or even null. However, these enhanced, situation-aware proposals rely on ontology-like

techniques that are not fine grained enough to detect light, but relevant, tendencies on supplier’s behaviour.

In this paper, we propose a technique that allows the extraction of positive and negative tendencies of

suppliers in the fulfilment of established contracts. This technique can be used with any of the existing

“traditional” CTR systems, improving their ability in selectively selecting a partner based on the

characteristics of the situation in evaluation. In this paper, we test our proposal using an aggregation engine

that embeds important properties of the dynamics of trust building.

1 INTRODUCTION

Several technologies are being studied and applied

in the general process of computerized supply chain

management. Computational trust management is

one such technology that will allow extending

electronic sourcing to world-wide located, non

registered and probably unknown business partners.

With this technology, a business entity will be able

to search the suppliers offer space and to filter the

ones that are fitted to the entity current needs, in a

scale of the size of the Internet.

The first generation of CTR systems addressed

the representation and the aggregation of trust

evidences into trustworthiness scores for evaluating

trustees, and most of these proposals are based on

some sort of statistical aggregation methods (e.g.

Ramchurn, Sierra, Godo and Jennings (2004),

Sabater (2003), Jøsang and Ismail (2002), Zacharia

and Maes (2000), Erete, Ferguson and Sen (2008),

and Huynh, Jennings and Shadbolt (2006)). Other

works proposed more sophisticated engines that

considers the dynamics of trust in the computation

of confidence scores, in theoretical and practical

terms (e.g. Elofson (1998), Falcone and

Castelfranchi (1998), Jonker and Treur (1999),

Marsh and Briggs (2008), and Melaye and

Demazeau (2005)). However, none of the current

computational trust approaches are mature enough to

be themselves trusted by real managers.

Trying to cope with this question, trust

community is moving towards a second generation

of models that explore the situation of the trust

assessment in order to improve its credibility, also

allowing for the estimation of trustworthiness values

when trust evidences on the trustee partner are

scarce or even null. However, few proposals have

been made on this specific area (see Tavakolifard

(2009), Neisse, Wegdam, Sinderen and Lenzini

(2009), Rehak, Gregor and Pechoucek (2006),

Fabregues and Madrenas-Ciurana (2009), and

Hermoso, Billhardt and Ossowski (2009)).

The purpose of this paper is two-folded. First, we

describe our proposal for an aggregation engine that

embeds three fundamental dynamics of trust

properties – asymmetry, maturity, and

distinguishable past –, and present our conclusions

about the relevance of the inclusion of such

properties in trust aggregation engines. Then, we

propose a situation-aware technique that allows the

Urbano J., Paula Rocha A. and Oliveira E. (2010).

REFINING THE TRUSTWORTHINESS ASSESSMENT OF SUPPLIERS THROUGH EXTRACTION OF STEREOTYPES.

In Proceedings of the 12th International Conference on Enterprise Information Systems - Artiﬁcial Intelligence and Decision Support Systems, pages

85-92

DOI: 10.5220/0002908600850092

 SciTePress

extraction of tendencies in the behaviour of agents.

This technique allows, for instance, to detect

whether a given supplier has a tendency to fail or to

succeed contracts that are similar to the current

business need (e.g. in terms of good, quantity and

delivery time conditions). We performed

experiments that show that this technique enhances

traditional CTR systems by bringing context into the

loop; i.e. it not only concerns if a given supplier is

generally trusted good or bad, but if it is trusted

good or bad in the specific contractual situation.

Also, this approach differs from other recently

proposed situation-aware proposals in the way that it

does not imply the use of hierarchical-based

structures (e.g. ontology) and is able to detect fine-

grain subtle dissimilarities in related situations.

Although we contextualized the use of our trust

system in the sourcing/procurement part of the

supply chain, agent-based trust and reputation

systems are of general interest in many other

domains (for instance, general business, psychology,

social simulation, system resources’ management,

etc), and apply to all social and business areas of the

society where trust is deemed of vital importance.

The remaining of this paper is structured as

follows: Section 2 describes our study about the

relevance of considering properties of the dynamics

of trust in the aggregation engine of CTR systems.

Section 3 describes the stereotype-based technique

we developed in order to complement traditional

CTR engines with situation-aware functionality.

Section 4 presents the experiments we run in order

to evaluate the proposed situation-aware technique,

and Section 5 concludes the paper.

2 USING TRUST DYNAMICS IN

THE AGGREGATION ENGINE

In previous work, we described an S-like

aggregation curve (see Figure 1) that allows for an

expressive representation of the dynamics of trust,

particularly, implementing the following properties:

 Asymmetry property, that stipulates that trust is

hard to gain and easy to lose;

 Maturity property, that measures the maturity

phase of the partner considering its

trustworthiness, where the slope of growth can

be different in different stages of the partner

trustworthiness;

 Distinguishably property, that distinguishes

between possible different patterns of past

behaviour.

The trustworthiness estimation of a given supplier

agent using this curve implies a slow growth upon

positive results when the partner is not yet trustable,

an acceleration when it is acquiring confidence, and

a slow decay when the partner is considered

trustable (i.e., in the top right third of the curve),

allowing for the definition of three different trust

maturity phases (the Maturity property). The

decrease movement upon negative results follows

the same logic, although the mathematical formula

subjacent to the curve includes parameter λ that

permits that trust grows slower and decays faster

(the Asymmetry property).

Figure 1: The S-like curve.

One can argue that we could use other S-like curves

instead of a sin-based one, such as the Sigmoid

curve. However, we intuitively feel that a Sigmoid

curve permits a probably too soft penalisation of

partners that proved to be trustable but that failed the

last n contracts. This can happens accidentally (e.g.

due to an unexpected shortage of good or to

distribution problems), but it is also described in the

literature as a typical behaviour of deceptive

provider agents, who tend to build up a trustworthy

image using simple contracts and then violate bigger

contracts exploring the acquired trustworthiness.

2.1 Evaluation of Trust Properties

Previous work provides a detailed description of the

S-like curve, as well as an experimental evaluation

of its behaviour. In this section, we summarize the

main conclusions we obtained when we

experimentally compared it (thereafter called the S

approach) to a weighted mean by recency approach

(that we named WMEAN), a common approach

seen in literature for traditional CTR aggregating

engines (cf. Huynh, Jennings and Shadbolt (2006)).

In this work, we explored three different

scenarios. In the first scenario, we wanted to

compare the capacity of both approaches in

differentiating between different types of supplier

y(α)=δ.sinα +δ

=3π/2,

α=α +λ.ω

2



ICEIS 2010 - 12th International Conference on Enterprise Information Systems

agents, namely, the capacity of primarily choosing

‘good’ suppliers that with a high probability do not

violate a contract. In such a scenario, we observed

that the S approach outperforms the WMEAN

approach in its capacity of selecting ‘good’ partner

agents, in one hand, and in avoiding ‘bad’ partners,

in the other hand. One difference between both

approaches resides in the fact that in S all the

historical path is taken into account in the process of

trust construction, and partners have to accumulate

several good experiences in the past until they are

able to get an average to high trust score (the

maturity property). In opposition, the WMEAN

approach allows the selection of partners with fewer

past events. For instance, analyzing the traces of the

experiments, we verified that some bad choices of

WMEAN happened when the algorithm selected

partners with rather few contractual past evidences

(e.g. the pattern of the previous evidences to the time

of selection where V-F-F-V-F-F, where V means a

violated contract and F a fulfilled contract).

Another difference between the two approaches

is due to the asymmetry property of S. This seems to

be particularly important when identifying and

acting upon partners that show intermittent

behaviour (e.g. F-V-V-F-V-F-V-V-F-F-F-F-F). This

last pattern of behaviour is indeed severely punished

by the S approach, where violations weight more

than fulfillments (therefore penalizing undesirable

intermittent patterns), and where the last few

positive evidences are not sufficient to ‘push’ the

confidence level of the partner to the second third of

the curve.

In the second scenario, we intended to study how

S and WMEAN react in the presence of extreme

partners that have a bursty-like behaviour (i.e. that

switch between sequences of good and deceptive

behaviour). By analysis of the traces of the

experiments, we realized that both approaches act

quite differently as they tend to select different

partners in similar conditions. The main point to

consider here is that WMEAN, by privileging

recency, actually assigns high trust levels to

candidate partners that systematically behaved

deceptively in the past, had no classification for a

long time, and then got one positive classification in

the present. I.e., WMEAN-like approaches can

forgive too fast in certain temporal scenarios. One

could argue here that this forgiveness issue is solved

by increasing the size of the window used (i.e. the

number of the last past evidences considered);

however, in our experiments we found it hard to

select the optimal window size, as it deeply depends

on the frequency of the contracts (historical

evidences) made in the past. The forgiveness

question does not apply to S, due to the action of the

maturity property; however, we realized that S has a

somewhat bigger tendency to enter a burst of

deceptive behaviour and that it can be slower in

penalizing good partners immediately after they

invert their behaviour.

Finally, the last scenario intended to study the

abuse of prior information scenario defined in

(Zacharia and Maes, 2000), where ‘good’ partners

definitely invert their behaviour after a given

number of iterations. The results that we obtained

showed that S outperforms WMEAN in detecting

and penalizing the change of behaviour of originally

‘good’ partners, while WMEAN showed a

significantly higher tendency to choose ‘bad’

partners than S.

2.2 Remarks about the S Curve

Taking into account all the experiments performed,

we can conclude that the three properties of the

dynamics of trust embedded in S are effective in

distinguishing between different types of target

agents, therefore in detecting and acting upon

undesirable agents’ behaviours. Namely, the

asymmetry property penalizes intermittent

behaviour, the maturity property avoids selection of

partners who did not prove to be trustable enough,

and the distinguishable past property avoids the

phenomenon of forgiveness described above.

Considering this last property, we have a somewhat

different view than the one presented in (Sabater,

Paolucci and Conte, 2006), where the authors state

that the aggregation of evaluations shall not depend

on the order in which these evaluations are

aggregated.

In these experiments, we could not evaluate,

however, the potential full benefits of the curve

shape against simpler curves that do show similar

trust dynamics properties (e.g. curves with linear

shape). In fact, as stated previously, S considers

different growth/decay slopes in different stages of

the trustworthiness acquisition of a target agent, and

it also presents a sigmoid-like shape. The choice of

this shape was based on the concept of the hysteresis

of trust and betrayal, from Straker (2008). In this

work, the author proposes a path in the form of a

hysteresis curve where trust and betrayal happens in

the balance between the trustworthiness of a self and

the trust placed on the self. The S curve simplifies

the hysteresis approach by using just one curve for

both trust and betrayal representation and

considering three different growth/decay stages:

REFINING THE TRUSTWORTHINESS ASSESSMENT OF SUPPLIERS THROUGH EXTRACTION OF

STEREOTYPES

Creating Trust (first third of the curve), Trust is

Given (second third of the curve), and Taking

Advantage (last third of the curve).

Performance tests of the S representation against

a simpler curve were performed. This new simpler

curve uses λ and ω parameters from S (cf. Figure 1)

to update the trustworthiness value of target agents,

but it lacks the softness round curve at Creating

Trust and Taking Advantage extremes. The results

of these experiments show similar performance of

both curves in the tested scenarios. Therefore, we

conclude that we need different, much more

complex models of target population to further study

the impact of the sigmoid-like shape of S on its

capability of distinguishing between partners. We

leave this topic for future work.

3 THE PROPOSED SYSTEM

3.1 Motivation for Situational Trust

Computational trust estimations help the trustier

agent to predict how well a given candidate partner

will execute a task and to compare between several

candidate partners. However, there are some

questions that a real-world manager would pose

before making a decision that cannot be answered by

simply aggregating available trust evidences into

trust and reputation values. These questions involve

somehow a certain level of intuition. We propose to

first analyze three scenarios that might occur in real

world business and that would help to understand

this concept.

In the first scenario, an agent may decide to

exclude from selection a candidate partner with

which it had never entailed business before but that

it knows that rarely fails a contract, just because the

agent intuitively fears that this partnership would not

be successful. For example, a high tech company

may fear to select a partner from a country of origin

without high technology tradition, even though this

partner has proved high quality work in the desired

task in the recent past. We call this situation the

intuitive fear. For this scenario, it would be desirable

that the selector agent could reason taking into

account additional contextual information about the

characteristics of the entity represented by the

candidate agent. For instance, the presence of key

figures such as the annual turnover or the number of

employees of the entity would allow the selector

agent to better know the entity. Also, the

establishment of argumentation between both parties

is a real-world procedure that could be automated

into the computational decision process. We address

the intuitive fear situation in future work.

In the second scenario, the agent may decide to

exclude from selection a candidate partner that is

currently entering the business, for which there is

not trust and/or reputation information yet. This

scenario deals with the problem of newcomers, for

which there is no information about prior

performance, and we name it absence of knowledge.

(Huynh et al., 2006) suggest that in these cases the

use of recommendations and institutional roles could

be useful to start considering newcomers in the

selection process. Although we do not address this

situation in this paper, we propose here to use

conceptual clustering of entities’ characteristics in

order to generate profiles of business entities. In a

second step, the profile of the newcomer is

compared with the profiles of business entities for

which there is some trust information and an

estimation of the newcomer trustworthiness is

inferred. This approach implies that the

characteristics of the business entities are available,

which is a reasonable assumption for centralized

virtual market places and virtual organizations built

upon electronic institutions, and might also be

applied to more decentralized approaches by

transmission of this kind of entities’ knowledge

between communicating agents.

Finally, in the third scenario, the selector entity

knows that a candidate partner is well reputed in

fulfilling agreements in a given role and context

(e.g. selling cotton zippers to European countries),

but it is afraid that the candidate is not able to

provide high quantities of the material in a short

period of time. We name this situation the contextual

ignorance. In this scenario, the evaluator agent

knows that the candidate partner is trustworthy in a

particular business scenario, or even that is generally

trustworthy, but needs to know how well it would

adapt to a different type of business. In this section,

we address this question by presenting a description

of our situation-aware technique, a component

complementary to the CTR aggregation engine that

is intended to give extra information to the trustier

agent by computing a value of how well the

candidate partner fits in the selector current needs,

as defined in the issued call for proposals (CFP).

3.2 The Situation-aware Technique

Every time a client issues a CFP, it may receive

several proposals from suppliers. In order to select

the best proposal, the client (trustier) computes a

general trustworthiness score for each supplier/

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

proposal. This score evaluates not only the general

behaviour of the supplier (trustee), but also the

adequacy of its business behavioural profile to the

CFP in question. The algorithm for computing the

trustworthiness value of a trustee is given next:

computeTrust (CFP, evds)

begin

negEvds ← getNegEvds (evds)

strtype ← getStrtype (negEvds)

trustScr ← compare (CFP, strtype)

if trustScr is 1

trustScr ← computeTrust (evds)

end

The algorithm above starts by evaluating the fitness

of the received proposal in relation to current CFP.

At line 3, all the evidences of the trustee that have a

negative outcome are put in one class. This negative

outcome can represent, for example, the past

transactions of the trustee that triggered relevant

contractual sanctions, although the meaning of such

outcome can be established by each individual client

agent. Then, at line 4, a stereotype is extracted for

this class, which means that the most significant

contractual characteristics of the evidences in the

class are going to be extracted. Depending on the

degree of the required extent of frequency increase

(cf. parameter α in equation 1) and on the evidence

set of the trustee, it is possible that the algorithm

does not return any stereotype. We must refer here

that this is an online process that is repeated every

time a new trust assessment is performed, which

allows to capture the variability of the behavior of

trustee agents at any time.

At line 5, the stereotype extracted (if any) is

compared to the current business need (CFP). A

match between stereotype and CFP attributes means

that the supplier/trustee has a tendency to fail this

type of contracts, and therefore the comparison

function returns a zero value, that would be the final

trustworthiness score for the trustee proposal.

Otherwise, there is no evident signal that the

supplier is inapt to perform the current transaction,

and its final trustworthiness score is computed using

the S approach described previously, or any other

CTR ‘traditional’ system (lines 6-7).

As can be seen above, we simplified the

proposed situation-aware technique by using (by

now) just a negative class of the evidences of the

trustee. The use of the positive class and the use of

distinct degrees of fitness will allow refining our

algorithm and this constitutes ongoing work.

We further describe the management of

stereotypes in the next sections. First, in order to

clarify the overall process, we describe how

contractual information (CFP and trustee evidences)

is represented in current implementation.

3.2.1 Representation of Information

In current implementation, contractual information

is represented by the tuple <A

, A

, At

..At

, t, o>,

where:

 Ac

∈

C is an agent from the set C of clients’

agents (i.e., the trustier agent);

 As

∈

S is an agent from the set S of suppliers’

agents (i.e., the trustee agent);

 At

∈

AT is the value of an attribute from the set

AT of n contract attributes (e.g. good, quantity

and delivery time);

 t is the timestamp of the transaction. Although

we are not using this value in current

implementation, it is needed in aggregation

systems that weights evidences by their recency;

 o

∈

{T, F} is the outcome of the contract, either

representing successful (true) or violated (false)

contracts by the supplier.

For instances, the evidence <A

, A

, cotton, 360000,

7, t, false> means that agent A

contractualized with

at time t the acquisition of 360000 meters of

cotton to be delivered within 7 days, and that A

failed to deliver the product in the aforementioned

conditions.

We must refer here that prior to stereotype

extraction, all evidence attributes are quantified to

categories or quantitative values using a fuzzy

approach. In current implementation, both quantity

and delivery time values are quantified to low,

medium and high categories.

Concerning the representation of trust, we

represent the trustworthiness score of trustee agent

relative to the current business situation as trust

, context

)

∈

[0, 1], where context

is an instance

of the context space Ctx constituted by all possible

combinations of the fuzzyfied values of the

attributes in set AT.

3.2.2 Stereotype Management

The extraction of stereotypes from the evidences

contained in each class is done using the metric of

equation 1 that measures the increase in the

frequency of an attribute within the community. This

metric was proposed in (Paliouras,

Karkaletsis, Papatheodorou and Pyropoulos, 1999).



#

#







#

#





(1)

REFINING THE TRUSTWORTHINESS ASSESSMENT OF SUPPLIERS THROUGH EXTRACTION OF

STEREOTYPES

In the equation above, #InstAttClass is the number

of times that a given attribute appears in the class,

#InstClass is the total number of evidences in the

class, #InstAttTotal is the number of times that the

attribute appears in all classes, and #InstTotal is the

total number of evidences kept for the trustee. As

mentioned before, parameter α is the degree of the

required extent of frequency increase, and

determines the granularity of stereotype extraction.

Finally, the comparison between a stereotype and

the current CFP is done attribute by attribute. Figure

2 illustrates a CFP request <A

, ?, chiffon, 1080000,

7, t, ?> from agent A

, whose quantity and delivery

time values are quantified into high and low values,

respectively, and the negative stereotype extracted

for an hypothetic proposal of agent X. The

stereotype means that, whatever fabric and quantity

is considered, agent X has a tendency to fail

contracts with low delivery time. Therefore, a match

is detected and the trustworthiness score of X for the

current proposal is set to zero.

Figure 2: Examples of a CFP and a stereotype.

4 EXPERIMENTS

In order to evaluate the benefits of the proposed

situation-aware technique (hereafter named SAT),

we run a series of experiments where a traditional

CTR aggregation engine – represented by the S

approach – was compared to the global solution

constituted by both S and SAT components.

4.1 Experimental Testbed and

Methodology

We run all the experiments in the Repast tool

(http://repast.sourceforge.net). The experiments

simulated a virtual textile marketplace, where at

every round every client agents post buying leads (in

the form of call for proposals) discriminating a

fabric to buy and correspondent quantity and

delivery time, and supplier agents propose in

response to these leads if they have the described

quantity of the fabric. Table 1 presents the

configuration options for the experiments.

In these experiments, we wanted to evaluate if the

situation-aware technique would improve the ability

of the trust

Table 1: Configuration of experiments.

Fabrics {Chiffon, Cotton; Voile}

Quantities {Low, Medium, High}

Delivery Time {Low, Medium, Big}

# buyers 20

# of sellers 50

Types of sellers Chosen upon a uniform

distribution over the types

{“S

”, “S

HDT

”, “S

HFB

”}

# rounds / # runs 100 / 40

α threshold 0.25

system in selecting partners taking into account the

current business needs. Therefore, we run the same

experiment using, first, just the S component, and

then the global solution of S plus SAT. We used the

utility criterion to compare both approaches: in each

round, the utility of a client agent was 1 if the

contract done in this round is successful and 0 if the

contract was violated. Therefore, the best approach

is the one that gets the higher average utility of all

clients in all rounds, i.e. the one that is more

efficient in selecting the best partners for every CFP

attributes at any time.

We also used a specific population of suppliers

constituted by three different types of suppliers, each

one showing some kind of handicap in fulfilling a

contract, as shown in Table 2.

Table 2: Different types of Suppliers.

Type Description

HQT

Probabilistically succeeds 95% of the

established contracts, except the ones that

involve the delivery of high quantities,

which probabilistic fails 95% of the time

HDT

Probabilistically succeeds 95% of the

established contracts, except the ones where

the delivery time is low, which probabilistic

fails 95% of the time

HFB

Probabilistically succeeds 95% of the

established contracts, except the ones that

involves the delivery of a given fabric,

which probabilistic fails 95% of the time

For example, a S

HQT

supplier would have a handicap

in providing the service if the quantity to provide is

high. Therefore, the best approach in evaluation is

the one that is more capable of detecting, and

reacting to, these types of handicaps.

4.2 Results

In every experiment, we measured the number of

successful contracts per type of target agents and per

approach, and averaged this number over the total

Stereot

e: Agent X, null, null, low, false

CFP: chiffon, 1080000, 7

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

number of rounds. In the best case possible, each

client is able to identify the handicap of every

provider and to select the best proposal, leading to

an average of 95% of successful contracts (Table 2).

The results obtained show that the SAT approach

gets, in average, 85.21% of utility, which means that

it is less than 10% away from the theoretical best

result. Also, the traditional approach gets in average

77.82% of utility, performing relevantly poorly than

its situation-aware counterpart.

Figure 3 shows the average number of successful

contracts per round, including the trendlines for the

traditional approach (above) and for the situation-

aware approach (below).

Figure 3: Average number of successful contracts per

round for S (above) and SAT (below) approaches.

We can observe from the figures above that,

although simple, the SAT algorithm is able to extract

correct stereotypes for each trustee agent with a few

number of past contractual evidences, for the

experimented population. This is a very important

issue in several domains, such as in the textile

industry, where direct or even indirect evaluations of

a given supplier might be scarce.

Also, by analyzing the traces of the experiments,

we observed that the learning curve for the situation-

aware solution is consistently more evident than the

one of the traditional approach. Figure 4 plots the

relative number of unsuccessful contracts per type of

supplier for S (above) and SAT (below), obtained in

one run of the experiments. From several of these

observations, we verified that, no matter what the

first choices were concerning the initial selection of

partners, the situation-aware solution often succeeds

in stabilizing the selection of all three types of

suppliers with low values of violated contracts.

On the other hand, the traditional approach keeps

selecting suppliers based only on the aggregate trust

Figure 4: Relative number of unsuccessful contracts per

supplier type without (above) and with (below) the

situation-aware technique.

score. As the three types of suppliers have equal

probability of failing (and therefore approximate

trustworthiness) if the analysis of the context is

excluded, they can be equally chosen for a given

CFP independently of their specific handicap.

In fact, looking at Figure 4, we observe that

suppliers with a handicap on quantity suffered from

a cold start, most probably because they were

initially selected to provide high quantities of

material. As S is not able to capture the handicaps –

and as suppliers of this type would tend to succeed

on all other contracts they are engaged to, therefore

maintaining some level of trustworthiness – the

algorithm will continue to select suppliers with

quantity handicap to provide high quantities of

material.

4.3 Interpretation of the Results

The results obtained show that the traditional

approach of aggregating trust, even when enhanced

with heuristics based on the dynamics of trust,

presents limitations when the suppliers in evaluation

present tendencies of failure. In these situations, the

extraction of behaviour stereotypes is effective in

discriminating the best suppliers to be chosen for

any particular business need. Also, this technique

showed to be effective since the first rounds of the

experiments, meaning that it is adequate to situations

where the available number of trust evidences about

the supplier in evaluation is scarce.

REFINING THE TRUSTWORTHINESS ASSESSMENT OF SUPPLIERS THROUGH EXTRACTION OF

STEREOTYPES

5 CONCLUSIONS

In this paper, we presented a simple situation-aware

technique (SAT) based on the extraction of

stereotypes of agents’ behaviour that can be used

with any traditional CTR system in order to enhance

the estimation of trustworthiness scores. Although

other situation-aware approaches are now being

proposed in the trust management field, the SAT

technique presents some benefits: i) it is simple and

can be used with any of the existing CTR

‘traditional’ aggregation engines; ii) it is an online

process, meaning that it captures the variability in

the trustee behaviour as it happens; iii) it does not

rely on ontology-based situation representation, and

therefore the extraction of the similarity between the

situation in assessment and the past evidences of

trustee agent does not require specific, domain-based

similarity functions; also, it allows for fine-grain

dissimilarity detection (e.g. it distinguishes between

the similar though different situations of providing

one container of cotton in 7 or in 14 days).

The SAT approach was evaluated using a

traditional aggregation engine approach enhanced by

the inclusion of properties of the dynamics of trust.

Although these properties showed to be beneficial,

we conclude that the study of the benefits of a

sinusoidal like shape that follows Straker (2008)

work on the area of Psychology needs proper

data/models concerning the behaviour of real-world

organizations; therefore, we will address the

acquisition of such data sets in future work.

ACKNOWLEDGEMENTS

The first author enjoys a PhD grant with reference

SFRH/BD/39070/2007 from the Portuguese

Fundação para a Ciência e a Tecnologia.

REFERENCES

Elofson, G., 1998. Developing Trust with Intelligent

Agents: An Exploratory Study. In Proceedings of the

First International Workshop on Trust, pp. 125-139.

Erete, I., Ferguson, E., Sen, S., 2008. Learning task-

specific trust decisions. In Procs. 7th Int. J. Conf. on

Autonomous Agents and Multiagent Systems, vol. 3.

Fabregues, A., Madrenas-Ciurana, J., 2009. SRM: a tool

for supplier performance. In AAMAS’09, 1375-1376.

Falcone, R., Castelfranchi, C., 1998. Principles of trust for

MAS: cognitive anatomy, social importance, and

quantification. In Procs. Int. Conference on Multi-

Agent Systems.

Hermoso, R., Billhardt, H., Ossowski, S., 2009. Dynamic

evolution of role taxonomies through

multidimensional clustering in multiagent

organizations. In Principles of Practice in Multi-Agent

Systems, vol. 5925, chapter 45, pp. 587–594.

Huynh, T. D., Jennings, N.R., Shadbolt, N.R., 2006. An

integrated trust and reputation model for open multi-

agent systems. In Autonomous Agents and Multi-Agent

Systems, Vol. 13, N. 2, September 2006, pp. 119–15.

Jonker, C. M., Treur, J., 1999. Formal Analysis of Models

for the Dynamics of Trust Based on Experiences. In

Procs. of the 9th European Workshop on Modelling

Autonomous Agents in Multi-Agent World: Multiagent

System Engineering. F. J. Garijo and M. Boman.

LNCS, vol. 1647. Springer-Verlag, London, 221-231.

Jøsang, A., Ismail, R., 2002. The Beta Reputation System.

In Proceedings of the 15th Bled Electronic Commerce

Conference, Sloven.

Marsh, S., Briggs, P., 2008. Examining Trust, Forgiveness

and Regret as Computational Concepts. Computing

with Social Trust. Springer, ed. J. Golbeck, pp. 9-43

Melaye, D., Demazeau, Y., 2005. Bayesian Dynamic Trust

Model. CEEMAS 2005: 480-489.

Neisse, R., Wegdam, M., Sinderen, M., Lenzini, G., 2009.

Trust management model and architecture for context-

aware service platforms. In On the Move to

Meaningful Internet Systems. LNCS, pp. 1803-1820.

Paliouras, G., Karkaletsis V., Papatheodorou, C.,

Pyropoulos, C. D., 1999. Exploiting Learning

Techniques for the Acquisition of User Stereotypes

and Communities. In Procs. of UM99.

Ramchurn, S., Sierra, C., Godo, L., Jennings, N.R., 2004.

Devising a trust model for multi-agent interactions

using confidence and reputation. In Int. J. Applied

Artificial Intelligence (18) 833–852.

Rehak, M., Gregor, M., Pechoucek, M., 2006.

Multidimensional context representations for

situational trust. In IEEE Workshop on Distributed

Intelligent Systems: Collective Intelligence and Its

Applications, pp. 315-320.

Sabater, J., 2003. Trust and Reputation for Agent

Societies. Number 20 in Monografies de l’institut

d’investigacio en intelligència artificial. IIIA-CSIC.

Sabater, J., Paolucci, M., Conte, R., 2006. Repage:

Reputation and image among limited autonomous

partners. In Journal of artificial societies and social

simulation, Vol. 9, pp. 3.

Straker, D., 2008. Changing Minds: in Detail

. Syque

Press.

Tavakolifard, M., 2009. Situation-aware trust

management. In RecSys '09: Proceedings of the third

ACM conference on Recommender systems, 413-416.

Zacharia, G., Maes, P., 2000. Trust management through

reputation mechanisms. In Applied Artificial

Intelligence, 14(9), 881–908.

ICEIS 2010 - 12th International Conference on Enterprise Information Systems