TRUST MODEL FOR HIGH QUALITY RECOMMENDATION

G. Lenzini, N. Sahli and H. Eertink

Telematica Instituut, Brouwerijstraat 1, 7523 XC, Enschede, The Netherlands

Keywords:

Trust Model, Recommendation, Argumentation, Agent.

Abstract:

In this paper, we propose a trust management model for decentralised systems that improves the quality of rec-

ommendations that members of a virtual community get about the trustworthiness of objects. In our system, as

in well-known solutions, members of the community evaluate (i) the functional trust in an item by the analysis

of the object’s qualities, past experience, and recommendations and (ii) the referral trust in a recommender by

the analysis of the recommender’s qualities and reputation based on personal experience. Moreover, in our

trust model, each principal debates with its recommenders about the justiﬁcations given to support a recom-

mendation. Thus, the usefulness and the reliability of a recommendation depend also on the strength of the

arguments supporting the recommendation. A measure of this strength results after the member has played an

argumentation game with the recommender. Therefore, the recommendations that are taken into account are

those which better match the member’s proﬁle and way of reasoning. Our trustworthiness evaluation algorithm

is context dependent and able to collect both direct and indirect information about trustees. Our trust model

is part of an agent-based architecture we propose for decentralised virtual communities. This architecture

provides our system with autonomy, unobtrusiveness, user mobility, and context-awareness.

1 INTRODUCTION

A virtual community (also e-community or on-line

community) is a group of socially interacting peo-

ple whose interaction is supported by computer sys-

tems (Preece, 2000). People interact primarily with

e-mails, chats, etc. rather than face to face. One

common application of virtual communities is sharing

opinions about objects of interest, for example rec-

ommendations and ratings. By using opinions, each

member can take a shortcut to items she/he likes with-

out having to try them or to experience many similar

items. Unsurprisingly, this facility has become popu-

lar on the Internet. “Amazon”, for example, rates each

item with stars. This overall rating is obtained by av-

eraging the ratings provided by users on the quality

of the product as a whole. Any member can write a

review. In IENS.nl, a Dutch site for rating restaurants,

and apart from the overall rating of a given restaurant,

a user can also rate/consult a reﬁned set of criteria re-

lated to the restaurant such as quality of food, quality

of service, decor, and price. Such systems (i.e., Ama-

zon, IENS, e-Bay) are mainly centralised; they rely on

all peers’ ratings and give the same recommendations

to all peers. In these systems, the inference process

that leads to a recommendation (either in the form of

a rating or a suggestion) is usually hidden to the user.

It is not possible to perform a qualitative analysis on

the reliability of the suggestion provided nor on the

process that has been used to compose it. It is also

not possible to personalise the ratings w.r.t. the user’s

way of reasoning. As a consequence, a user who is

looking for a recommendation which ﬁts his/her taste

must read a large number of reviews left by the other

users. In fact, she/he has to manually look for justiﬁ-

cations that support the choice which better matches

his/her proﬁle and expectations.

A decentralised approach has thus emerged as

a serious alternative (e.g., see (Sabater and Sierra,

2002; Miller et al., 2004; Teacy et al., 2005)). When

the trustworthiness and the personalisation of recom-

mendations is an issue of great concern, decentralisa-

tion might be more appropriate. In a decentralised ap-

proach the user only relies on recommendations sug-

gested by other peers in whom she/he trusts. We pro-

pose an agent-based decentralised system for trust-

based recommendations which goes farther than pre-

vious decentralised systems. Our system aims at im-

518

Lenzini G., Sahli N. and Eertink H. (2008).

TRUST MODEL FOR HIGH QUALITY RECOMMENDATION.

In Proceedings of the International Conference on Security and Cryptography, pages 518-528

DOI: 10.5220/0001921805180528

 SciTePress

proving the quality of recommendations. Since even

these recommenders could have different tastes and

different ways of judging, being trustworthy is not

enough to have suggestions accepted by the current

peer. Thus, we propose that the peer should also de-

bate with its recommenders about the arguments that

justify a recommendation in order to only select the

most likely appropriate (from this peer’s view point)

recommendations.

In this paper we focus on the argumentation pro-

cess and its impact on the trustworthiness of recom-

mendations. The remainder of the paper is organised

as follows. Section 2 gives an overview of the main

principles of our approach that are behind the qual-

ity of recommendations. Section 3 explains the rules

that model our trust evaluation system with focus on

the argumentation process. Section 4 explains how

to measure the “strength” of a recommendation. Sec-

tion 5 illustrates an example of the use of trust eval-

uation with argumentation. Section 6 presents the

agent-based architecture we propose to support our

trust model. Section 7 comments the related work and

Section 8 concludes the paper and points out the fu-

ture work.

2 OUR APPROACH

In (Lenzini et al., 2008) we presented a preliminary

solution to improve the quality of recommendations

in mobile and non-mobile open virtual communities.

A community is hosted in what we call a Virtual

Agora (in short, VA). A VA is a virtual open space

(e.g., web site, server) where active entities (e.g., con-

sumers) meet, interact, and share experiences about

items (e.g., goods, services) of interest. From this

point of view, many services already available match

the deﬁnition of a VA (e.g., the mentioned Amazon

and e-Bay). Moreover, a VA satisﬁes also the follow-

ing characteristics: (i) openness, entities from various

sources can freely join or leave at any time; (ii) decen-

tralisation, no central authority controls entities (but

we admit a certain centralised facility to search for

members, see Section 6), and (iii) persistence, enti-

ties (if desired) can be continuously available. In this

paper, we present an advanced solution for the im-

provement of the quality of recommendations within

a VA. It is based on the following three ideas:

(1) In the VA, each user is represented by a software

agent, called delegate agent, which behaves on

behalf of its user (even when the user is off-line)

(2) Each delegate agent maintains a personal regis-

ter of rated items (Register of Rated Items) and a

personal trust-weighted register of recommenders

(Register of (Un)Trusted Recommenders).

(3) When evaluating the trustworthiness of an item,

a delegate agent accepts only high quality recom-

mendations. The quality of a recommendation de-

pends upon both the referral trust of the source

(from this agent’s point of view) and the measure

of the “strength” of the arguments that support the

recommendation.

We now comment the three aforementioned ideas.

Point (1) concerns the architectural design of a

VA, which is part but not the main goal of this pa-

per. Brieﬂy, peers need autonomy to be able to act on

behalf of their users. To design an effective VA which

fulﬁls this requirement, we chose an agent-based ar-

chitecture. Each peer is represented by a delegate

agent that once moved to the VA is constantly inter-

acting with the other delegate agents and updating its

user’s registers of trust. More details of the architec-

ture are given in Section 6.

Point (2) concerns the two trust-weighted regis-

ters that each delegate agent uses to evaluate the func-

tional trust on an item and the referral trust on a rec-

ommender. In the register called Register of Rated

Items (in short Trat), an agent keeps trace of the func-

tional trust in the items that it has evaluated so far,

directly (from personal experience), indirectly (re-

ported by other recommenders) or both. In the regis-

ter called Register of (Un)Trusted Recommenders (in

short Trec) an agent keeps trace of its referral trust on

other peers in giving recommendations. The two reg-

isters are initiated and updated according to rules that

are explained in Section 3.

Finally, point (3) regards the quality of recommen-

dation. When a peer (i.e., delegate agent) asks for

recommendations it evaluates the information it re-

ceives according to the referral trust of the source and

to the “appropriateness” of the recommendations with

respect to its proﬁle, taste, and way or reasoning. The

delegate agent reaches this latter goal by “discussing”

with its recommenders about the arguments that they

provide to justify their recommendations. Eventually,

whether or not the agent accepts a recommendation,

depends upon both the warrant of the recommenda-

tion and the referral trust of the recommender. Thus,

“warranted” recommendations coming from not so

trustworthy members may be accepted, whereas “un-

warranted” recommendations coming from trustwor-

thy recommenders might be discarded. We believe

that an opinion built upon a small number of recom-

mendations of high quality is generally more useful

than a digest of anonymous, often contradicting, and

potentially unjustiﬁed recommendations.

TRUST MODEL FOR HIGH QUALITY RECOMMENDATION

519

3 TRUST RELATIONSHIPS

MANAGEMENT

This section explains our trust model and the algo-

rithm that a peer uses to build and to update its Trat

and Trec. The basic rules have already been presented

in (Lenzini et al., 2008). Here, we introduce and study

new rules whose goal is to quantify the quality of a

recommendation in the trust evaluation process. We

also present an implementation based on argumenta-

tion theory (e.g., see (Prakken, 2006)). We introduce

and discuss the concept of argumentation trust, which

we use as a measure of the recommendation strength.

To our best knowledge, the framework we present is

innovative and deﬁnitely more dynamic than our pre-

vious solution.

The attention to the justiﬁcation of a recommenda-

tion comes from the need of distinguishing between a

measure of the reputation of a recommender (in giv-

ing honest recommendations) and the appropriateness

of a speciﬁc recommendation in a given context. For

example, even if peer B is renowned to be an expert in

restaurants (fact that she/he has proved many times),

her/his recommendation about going to the Italian

restaurant “La Barcarola” may be discarded by a prin-

cipal A if A and B have different opinions about what a

good Italian menu should be. What is excellent from

B’s view point could be inappropriate for A (which

has, in this particular context, different preferences).

On the other hand, A may accept the recommendation

of another peer D which is less qualiﬁed than B but

whose justiﬁcation about going to the “L’Ostricaro”,

another Italian restaurant, better convinces A. Later,

when A effectively tries this restaurant and likes it, D

also sees its trustworthiness growing from A’s point of

view; D’s recommendation has perfectly satisﬁed A’s

expectations. It is worth noticing that B’s reputation

remains intact. A has no interest in saying that B is

a bad recommender only because A has not followed

B’s advice. It is not B’s reputation that in in ques-

tion here, but rather the justiﬁcation that B gave about

her/his recommendation, which didn’t match A’s way

of reasoning from the beginning.

The Register of Rated Items of a member A (writ-

ten Trat(A)) is a set of trust relationships, written as

(i,m)

−−→

G;C

Here, A is an agent and b is an item. All b’s that

are in Trat(A) have been evaluated by A directly, in-

directly, or both. The value m is a measure of the

functional trust (positive or negative) that A has in b.

The natural number i is a time-stamp from which to

deduce the age of the trust relation. The subscript G

is the set of recommendations and of recommenders

that have been consulted in composing m. The sub-

script C is the context in which the trust relation takes

or has taken place. This parameter makes the trust re-

lationship context-dependent. If b is a restaurant, for

example, one’s evaluation may change depending on

the social context of the user (e.g., with colleagues,

with partner, or with family). The subscript σ stands

for a trust aspect (called trust scope in (Jøsang et al.,

2006)). It represents the purpose of the trust relation.

In Trat(A) there is a multiplicity of trust relationships

between A and b for each aspect among σ

, σ

, etc.

To simplify our exposition, we consider only one σ

which we omit from our notation. Anyhow, it must

be clear that the overall trust relationship between A

and b depends upon all the trust aspects.

The Register of (Un)Trusted Recommenders of a

member A (called Trec(A)) contains trust relation-

ships written as

(i,ω)

==⇒

Here A and D are both agents. If a member D is in

Trec(A), this means that D’s trustworthiness has been

evaluated by A. The natural number i is a time-stamp,

and the value ω represents an amount of referral trust

that A has in D. Set O carries information about the

items that D has so far recommended to A and the

relative recommendations.

3.1 Trust Evaluation Algorithm

The process of trust evaluation is formalised with an

inference system, which is part of an agent A’s intel-

ligence. Figure 1 and 2 summarise the rules of the

inference system used to build Trat(A) and Trec(A),

respectively. Each trust evaluation rule, in A’s intelli-

gence, has form:

premises

conclusion

A conditions Premises are re-

quired in order to infer the conclusion. The inference

mechanism is processed if conditions (if present) are

fulﬁlled. The subscript, here a generic agent A, indi-

cates that this rule is applied by A. We will omit the

subscript in the rest of the paper since we present all

rules from A’s point of view.

As a matter of computation, we use two categories

of arrows, namely temporary (or tentative), written

•→ and •⇒ resp., and eventual (or conclusive), writ-

ten → and ⇒ respectively (see Figures 1 and 2). A

temporary trust relation emerges from an incomplete

trust evaluation; additional inputs can still affect the

trust value. An eventual trust relation indicates that

the trust evaluation has ended; no more input can

change the trust value for time i. The registers Trat(A)

and Trec(A) are composed of only eventual trust re-

lationships. In the following, we assume an algebra

SECRYPT 2008 - International Conference on Security and Cryptography

520

Bootstrapping (1)

(a)

[eval

({b

},C ) = m]

A•

(i,m)

−−−−−−→

{(A

,m)};C

(c)

[tag

(b,C ) = m]

(i,m)

−−−−−−→

{(A,m)};C

(b)

[eval

({b

∼b},C ) = m]

A•

(i,m)

−−−−−−→

{(A

,m)};C

Recommendation (2)

(a)

(i,ω)

==⇒

−

D D

(i−1,m)

−−−−−−→

{(D,m)};C

b A

i,ω

(−−

−−+

A•

(i,ω

⊗m)

−−−−−−→

{(D,m)};C

i > 0

ω ⊕ ω

= ω

> ω

(b)

(i,ω)

==⇒

−

D D

(i−1,m)

−−−−−−→

{(D,m)};C

i,ω

(−−

−−+

A•

(i,ω

⊗(ω

(d(b

,b))

⊗ m))

−−−−−−−−−−−−−→

{(D,m)};C

∼ b, b

6= b

i > 0

ω ⊕ ω

= ω

> ω

(c)

(i,ω)

==⇒

−

D D

(i−1,m)

−−−−−−→

{(D,m)};C

b A

i,ω

(−−

−−+

A•

(i,ω

⊗(ω

(d(C

,C ))

⊗ m))

−−−−−−−−−−−−−−→

{(D,m)};C

∼ C , C

6= C

i > 0

ω ⊕ ω

= ω

> ω

(d)

(i,ω)

==⇒

−

D D

(i−1,m)

−−−−→

G;C

b A

i,ω

(−−

−−+

A•

(i,ω

⊗m)

−−−−−−→

{(D,m)};C

G 6= {(D,m)}

A 6∈ G

↓1

i > 0

ω ⊕ ω

= ω

> ω

Experience (3)

(a)

(k,m)

−−−−−−→

{(A,m)};C

A•

(i,ω⊗m)

−−−−−−→

{(A

,m)};C

ω = ω

i−k

k < i

(b)

(k,m)

−−−−−−→

{(A,m)};C

A•

(i,ω⊗(ω

⊗m))

−−−−−−−−−→

{(A

,m)};C

∼ b

b 6= b

= ω

d(b

,b)

ω = ω

i−k

k < i

(c)

(k,m)

−−−−−−→

{(A,m)};C

A•

(i,ω

i−k

⊗(ω

d(C

,C )

⊗m))

−−−−−−−−−−−−−−→

{(A

,m)};C

∼ C

C 6= C

k < i

Finalising (4)

(a)

A•

(i,m)

−−−→

G;C

b A•

(i,m

)

−−−−→

A•

(i,m⊕m

)

−−−−−−→

G∪G

↓1

∩ G

↓1

(b)

A•

(i,m)

−−−→

G;C

(i,m)

−−−→

G;C

Figure 1: Inference rules for the management of agent-item trust relationships. The recommendation rules, whose title is

boxed, and their implementation (in terms of argumentation) form the original contribution of this paper.

(TValues,⊕,⊗,0, 1) of trust values. The binary oper-

ator ⊕ is used to merge trust values, whilst ⊗ to dis-

count trust. Values 0 and 1 are the respective neutral

elements. The literature offers many computational

models that are in principle applicable.

Building and Updating Trat. Referring to Figure 1,

rules (1.a)-(1.b) allow A to apply its own procedure

eval

(x,y) (whose concrete implementation is left

unspeciﬁed at this level of abstraction) to estimate b’s

trustworthiness. The procedure takes as input the fol-

lowing two parameters: b

, which with a little abuse

of notation represents the information available about

b at time i and in context C . In case of lack of infor-

mation about b, a witness b

(i.e., b

∼ b) can be used

instead of b. The label {(A

,m)}; C in the resulting

trust relationship records that the value m has been

calculated at time i and in context C . Rule (1.c) mod-

els the action of rating b which follows a direct expe-

rience. A uses the function tag

(x,y) to tag b at time

i and in context C . The trust relation that emerges af-

ter a tagging is an eventual relation; there is no need

of other information to evaluate b’s trustworthiness.

Rules (2.a)-(2.d) in Figure 1 describe how we

manage recommendations. These rules extend and

renew the corresponding rules we gave in (Lenzini

et al., 2008), where we allowed an agent A to accept

a recommendation when the recommender’s referral

trust is above a certain threshold and when the justiﬁ-

cation given for the recommendation is accepted by A.

We had there no quantitative evaluation of the justiﬁ-

cation of a recommendation. In Figure 1, instead, the

trustworthiness of D’s recommendation is evaluated

TRUST MODEL FOR HIGH QUALITY RECOMMENDATION

521

Bootstrapping (5)

(a)

[estimate

) = ω]

A•

(i,ω)

==⇒

(b)

(i−1,ω)

====⇒

A•

(i,ω)

==⇒

i > 0

Update (6)

(a)

A•

(i,ω)

==⇒

D A

(i,m

)

−−−−−−−→

{(A

)};C

b A

(k,m

)

−−−−→

G;C

A•

(i,inc(ω))

======⇒

O∪{(b,k)}

k ≤ i,

(b,k) 6∈ O

({(D,m)} ∈ G)∧

∼ m)

(b)

A•

(i,ω)

==⇒

D A

(i,m

)

−−−−−−−→

{(A

)};C

b A

(k,m

)

−−−−→

G;C

A•

(i,dec(ω))

======⇒

O∪{(b,k)}

k ≤ i,

(b,k) 6∈ O

({(D,m)} ∈ G)∧

6∼ m)

Finalisation (7)

(a)

A•

(i,ω)

==⇒

D A•

(i,ω

)

===⇒

A•

(i,ω⊕ω

)

=====⇒

O∪O

O∩O

(b)

A•

(i,ω)

==⇒

(i,ω)

==⇒

Figure 2: Inference rules for managing agent-agent trust relationships.

by taking into account both the referral trust (that A

has already in D) and a measure of trust that emerges

from A’s analysis of the arguments that D gives to jus-

tify its recommendation. The relation A

i,ω

(−

−+

D indi-

cates that A’s evaluation of D’s recommendation (after

having discussed about its justiﬁcation) is ω. We call

this value argumentation trust Section 4 explains how

we estimate argumentation trust. Function ω ⊕ ω

re-

ﬂects a merged value between ω (argumentation trust)

and ω

(referral trust) in our algebra of trust values.

As a matter of example, it may happen that A accepts

a recommendation from a peer that is not so highly

trustworthy (e.g., a new member) but whose justiﬁ-

cation is strong. Similarly, a recommendation that

comes from a highly trustworthy recommender but

whose justiﬁcation is unconvincing, can be refused.

The resulting mechanism for (referral) trust evalua-

tion is more ﬂexible. The “strength” of justiﬁcations

(given by a recommender) is thus as important as the

reputation of the recommender.

Rules (3.a)-(3-c) show how to use the experience

of A about b, or about some of b’s witnesses (i.e.,

∼ b), to evaluate b’s temporary trust. Information

related to the same context, or to a compatible con-

text (i.e., C

∼ C ), can also be considered. When

witnesses or compatible contexts are used, the past

experience is discounted depending on a measure of

distance between b

and b or between C

and C , re-

spectively (see also (Toivonen et al., 2006)).

Rule (4.a) describes how the temporary evalua-

tions collected by different sources can be merged.

The condition G

↓1

∩ G

↓1

0 informally says that the

sources must be disjointed to avoid interferences. Fi-

nally, rule (4.b) ﬁnalises the trust process. The tempo-

rary trust evaluation becomes eventual. The recom-

menders, as well as their recommendations that have

been used in estimating b’s trustworthiness,are logged

in G. They are used to update Trec(A) (see next para-

graphs).

Building and Updating Trec. Referring to Fig-

ure 2, rule (5.a) initiates Trec(A). D’s proﬁle at time

i is evaluated by A with the function estimate

(x).

The value ω is an estimation of D’s referral trustwor-

thiness. Implementations of estimate

(x), out of

the scope of this paper, can be based on the amount

of “likelihood” between D and A for example, or on

the number of items that have been similarly ranked

by A and D. Rule (5.b) expresses that trust can be es-

timated by inheritance with a previous trust. Here we

do not model trust decay over time, or in other words

we do not discount a referral trust because of time.

Rule (6.a) and rule (6.b) say that A’s (referral)

trust in an agent D is increased (resp., decreased) for

each compatible (resp. incompatible) recommenda-

tion given by D. A recommendation is compatible

only in comparison with A’s personal experience (i.e.,

tagging). Here, overriding our notation, the symbol ∼

indicates the compatibility relation among opinions.

After A tags b, its trust in b is more based on facts

than on presumptions (according to our notation, this

is recognisable from the subscript {(A

,m)}). There-

fore, A is able to review the opinion she/he got via rec-

SECRYPT 2008 - International Conference on Security and Cryptography

522

ommendation and to modify her/his trust in the rec-

ommender accordingly. The comparison is context-

dependent. Rule (7.a) says that the temporary eval-

uations collected when updating the referral trust for

a recommender can be merged (operator ⊕) together.

A recommender might have given opinions on more

than one item and its overall referral trust depends

on the trust update for all of them. The condition

O ∩ O

0 ensures that the same update (per item)

is not considered twice. Finally, rule (7.b) ﬁnalises

the trust process. The temporary trust evaluation be-

comes eventual, and it will become part of Trat(A).

The items recommended along the way that con-

tributed to the referral trust are logged.

Implementation of the Model. To apply the rules

for Trat(A) and Trec(A) in an effective trust evalu-

ation algorithm, an order of preference among rules

is needed. Different ordering constraints reﬂect dif-

ferent attitudes in merging opinions and past expe-

riences. For example, an agent may prefer to ﬁrst

collect all opinions, then to merge them, and ﬁnally

to consider her/his past experience. In rules (3.a)-

(3.c) for example, different constraints over the time

variable k guide the search strategy in the past (e.g.,

choosing a maximal k implies considering the most

recent experience). Discussing these choices is out of

the focus of this paper.

4 ARGUMENTATION TRUST

This section explains how we measure the strength

of a recommendation given a set of arguments. We

ﬁrst remind the basic concepts of argumentation.In

logic and in artiﬁcial intelligence, argumentation is a

way of formalising common-sense reasoning. Cur-

rent implementations use both classical, monotonic,

logic and non-monotonic reasoning. Depending on

the implementation, an argument can be presented

as an inference tree, a deduction sequence, or a pair

(set of premises, conclusion). Most of the argumen-

tation systems can be characterised by ﬁve elements,

namely, (i) an underlying logical language, (ii) a con-

cept of argument, (iii) a concept of conﬂict between

arguments (which involves counter-arguments that at-

tack an original argument), (iv) a notion of defeat

among arguments, and (v) a notion of acceptability

of arguments according to a well-deﬁned criterion.

A counter-argument can attack an original argu-

ment directly, its premises, or the rules that have

been used to prove it. The right deﬁnition of attack

depends on the chosen implementation. In (Benta-

har and Meyer, 2007), which uses an implementation

based on classical logic, a counter-argument negates

an original argument or one of its premises. In (Garcia

and Simari, 2002), where a Defeasible Logic (Nute,

1994) implementation is used, counter-arguments are

supported by stronger proofs or stronger knowledge

than those supporting an original argument.

Attached arguments can be defended, for example

by producing a counter-argument against the counter-

argument. This interaction is expressed through and

argumentation protocol, where attacks and defences

interleave until the original argument is accepted, re-

fused, or declared undecided. These three concepts

will be used to quantify the strength of a recommen-

dation (see Deﬁnition 1).

The integration we present is general and inde-

pendent from the implementation chosen for the ar-

gumentation system.

Deﬁnition 1 (Argumentation Trust) . Let us sup-

pose two agents, A and D, and an argumentation

system Γ. A and D run N argumentation proto-

cols to debate over the arguments that D provides

to justify her/his recommendation. Let N

, N

+ N

= N) be the number of D’s initial ar-

guments that A accepts, refuses, and neither accepts

nor refuses (i.e., undeﬁned), respectively. Then the ar-

gumentation trust that A has on D is calculated with

the following function:

ω(N) =

+ N

+ (

)

(8)

where N

= N

+ N

is the number of arguments that

A has either accepted or not refused.

In (8), high trust values are reached with a high

number of not refused arguments. Thus, justiﬁcations

made of a large number of arguments can bring to

higher argumentation trust than justiﬁcations made of

a small number of arguments (see Figure 3).

Referring back to Figure 1, A

i,ω

(−

−+

D expresses that

A and D have run an argumentation protocol and, as a

result, A’s argumentation trust in D is ω(N) = ω.

5 RUNNING EXAMPLE

In this section we illustrate an example of trust evalu-

ation with argumentation. The argumentation game is

informally described. This means that this example is

not binded to a speciﬁc implementation. The instanti-

ation of argumentation into a formal scheme is left as

future work.

Let us assume that a member (Alice) of a slow

food community is looking for information about

TRUST MODEL FOR HIGH QUALITY RECOMMENDATION

523

0 2 4 6 8 10

Number of Arguments

0.2

0.4

0.6

0.8

Argumentation Trust

Figure 3: Graphic showing the argumentation trust function

when all, all but one, all but two arguments provided by D

are accepted by A, respectively (e.g., N

= 0,1,2 from top

to bottom). The more accepted arguments D provides the

higher its argumentation trust can be.

restaurant “La Bella Vita” (in short, b) in Amster-

dam. Alice wants to have a romantic dinner eating

Italian food. Her delegate agent A, which runs in

the Virtual Agora (i.e., the meeting infrastructure for

the slow food community), will collect information

about the trustworthiness of b. In the following, we

assume that the algebra of trust values is the real inter-

val [0,1] with a consensus operator x ⊕ y = (x + y)/2

and a discount operator x ⊗ y = x · y. Trust values in

[0,1] are informally read as follows: [0, 0.3] = distrust,

(0.3,0.5] = weak distrust, (0.5, 0.7] = weak trust, and

(0.7,1.0] = trust.

The goal of A is to evaluate the functional trust

of b at the current time (let us say w

). Formally, A

has to infer A

,m)

−−−−→

G;C

b, where C is the social context

“romantic dinner”. By consulting its register of rated

items Trat(A), A ﬁnds out that: (i) it has never rated

b before, but that (ii) it has evaluated restaurant “La

Bella Vita” in Utrecht (in short, b

) which belongs to

the same franchising as b. This rating was registered

when Alice was in Utrecht last week (week w

) for

a romantic dinner. The stored rating (functional trust

in b

) is 0.8. Formally, this means that A

,0.8)

−−−−−−−−→

{(A

,0.8)};C

∈ Trat(A).

Because b and b

belong to the same franchising

(i.e., b

∼ b), A estimates its trust in b by applying rule

(3.b). If we assume that the trust decay is ω

−w

1.0 (i.e., one week-old experience is fully signiﬁcant)

and the discount due to the use of a witness ω

d(b

,b)

0.94 (i.e., the trust experience with a restaurant of the

same chain is quite signiﬁcant), then A’s temporary

trust in b is 1.0 · (0.94 · 0.8) = 0.75. Formally:

,0.8)

−−−−−−−−→

{(A

,0.8)};C

A•

,0.75)

−−−−−−−−→

{(A

,0.75)};C

∼ b

−w

= 1.0

d(b

,b)

= 0.94

Agent A also looks for recommendation about

b. A ﬁnds out that (i) agent D, which is already in

Trec(A) and whose referral trust has been evaluated

0.8, recommends a rating 0.4 for b (i.e., according

to D, b is untrustworthy) (ii) agent E, which does

not belong to Trec(A) recommends a rating 0.8 for b

(i.e., E’s opinion is that b is a trustworthy restaurant).

Both agents mention the context “romantic dinner”.

By considering recommendations (i.e., rule (2.a)), A

can obtain two more instances A•

,m)

−−−−→

G;C

b. Later, A

can merge them with the instance obtained by Alice’s

past experience. First, A needs to evaluate its refer-

ral trust in D and E for the current time w

. Because

,0.8

===⇒

D ∈ Trec(A), A can “refresh” its referral trust

by applying rules (5.b) and (7.b), as follows:

,0.8)

====⇒

A•

,0.8)

====⇒

,0.8)

====⇒

Concerning E, A estimates its referral trust by ap-

plying rule (5.a) and later (7.b). If we assume that

estimate

(E) = 0.7, we have:

[estimate

(E) = 0.7]

A•

,0.7)

====⇒

,0.7)

====⇒

Without an analysis of justiﬁcation, and if we sup-

pose that A only accepts recommendations coming

from members whose referral trust is over the thresh-

old w

= 0.69, A would accept both D’s and E’s rec-

ommendations. By discounting and merging the two

recommendations (rules (2.a) and (4.a)), the trust in b

would be (0.8 · 0.4 + 0.7 · 0.8)/2 = 0.44. This result

would be merged, by applying the same rules, with

the trust that A gets from Alice’s past experience. This

would make a ﬁnal value of (0.75 + 0.44)/2 = 0.6.

Agent A would then ﬁnalise the trust evaluation pro-

cess (by rule 4.b). The ﬁnal conclusion is that b is a

weak trustworthy restaurant. Alice, who prefers trust-

worthy restaurants, would thus decide not to try b.

However, if argumentation is used, the ﬁnal judge-

ment could be different and more appropriate to Al-

ice’s proﬁle. In this case, A evaluates the argumenta-

tion trust of D by asking for a justiﬁcation about D’s

SECRYPT 2008 - International Conference on Security and Cryptography

524

low rating of 0.4. Different argumentation dialogues

take place between A and D, each for each aspects

that D has considered when composing its rating in b.

For example, the dialogue concerning the “ambiance”

informally appears like what follows:

A: why did you give 0.4?

(demand a justification)

D: I didn’t like the restaurant

because the restaurant was too quiet

(ground for a justification)

A: Yes, but this is romantic, isn’t it?

(counter-argument, demand a justification)

D: No, because there were too many moments

of silence between me and my partner

(ground for the justification)

A: I see, but this is not important for me

(rebuttal of argument)

Similar dialogues clarify the point of view of A and D

on two other aspects, namely, “quality of food” and

“price”. Let us suppose that only D’s justiﬁcation

about the “quality of food” is accepted by A. Thus,

according to formula (8), the argumentation trust be-

tween A and D is 1/(3 +

) = 0.3 and A

,0.3

(−−−

−−−+

When this value is “added” to the referral trust of D,

0.8, the resulting ω

is 0.55 (see rule (2.a)). Conse-

quently, D’s recommendation is not considered in this

tournament because it is below the threshold ω

On the contrary, after arguing with E, A accepts

all the arguments of E over “ambiance”, “quality of

food” and “price” (we do not show the dialogues

here). Agent A and E have a perfect match of opin-

ions, and the argumentation trust calculated with the

formula (8) is 3/(3 +

) = 0.9 and A

,0.9

(−−−

−−−+

E. When

this value is “added” to the referral trust of E, 0.7,

the resulting ω

= 0.7 ⊕ 0.9 is 0.8. The recommen-

dation of E is then accepted because it is above the

threshold ω

, and the temporary functional trust in b

is (0.9 · 0.8) = 0.72 according to rule (2.a) applied as

follows:

,0.7)

====⇒

E E

,0.8)

−−−−−→

(E,0.8);C

b A

,0.9

(−−−−

−−−−+

A•

,0.74

−−−−−→

(E,0.8);C

0.7 ⊕ 0.9 > 0.69

Finally, when this last value is merged with the trust

that A gets from Alice’s past experience, the ﬁnal trust

is (0.75 + 0.72)/2 = 0.74 (rule (4.a)):

A•

,0.75)

−−−−−−−−−→

{(A

,0.75)};C

b A•

,0.72)

−−−−−−−−→

{(E,0.72)};C

A•

,0.74)

−−−−−−−−−−−−−−→

{(A

,0.75),(E,0.72)};C

,0.74)

−−−−−−−−−−−−−−→

{(A

,0.75),(E,0.72)};C

In this case, A’s suggestion is that b is trustworthy and

Alice will try restaurant b.

After trying restaurant b, Alice rates it 0.9 (rule

(1.c); she liked the romantic ambiance of the restau-

rant. Consequently, agent A raises the referral trust

of E, for example, from 0.7 to 0.75 (see rule (6.a)).

In contrast, D’s referral trust remains unchanged, be-

cause its recommendation has not been considered by

A. D still can suggest nice restaurants later on.

6 AGENT-BASED SYSTEM

ARCHITECTURE

In order to implement the proposed trust model, our

delegate agents need the following capabilities (i)

Reasoning: it should be able to evaluate trust values,

build and update its knowledge (TRat and TRec), and

argue with other peers; (ii) Autonomy: it has to pro-

cess the aforementioned tasks autonomously (without

any manual assistance from its user); and (ii) Context-

awareness: it needs to capture the context of its user,

which is needed to reason about the trust.

To fulﬁl the aforementioned requirements, we de-

sign the VA as an open multi-agent system (open

MAS) (Barber and Kim, 2002), which represents a

scalable and ﬂexible system that matches our virtual

community concept. Moreover, the two main features

of open MAS members: (i) can freely join and leave

at any time and (ii) are owned by different stake-

holders with different aims and objectives, perfectly

ﬁt the description of our VA’s delegate agents. How-

ever, our architecture is not completely decentralised.

In fact, we use a central component called Bulletin

Board which is in charge of keeping up-to-date the

list of present members (useful for members discov-

ery) and the list of items to be evaluated (see also Fig-

ure 5).

With respect to the internal architecture of del-

egate agents, we adopt a Belief-Desire-Intention

(BDI) (Rao and Georgeff, 1995). The BDI model

offers an interesting framework to design delibera-

tive agents which are able to act and interact au-

tonomously and according to their mental states. The

main components of the delegate agent’s architecture

are illustrated in Figure 4. Rounded rectangles rep-

resent processes while rectangles represent the differ-

ent data. In the “Memory” component, two different

shapes are used to show whether the data is an in-

put (e.g., Proﬁle) or an output (e.g., Answer). In brief,

the “Goal Generator” (corresponding to Desires in the

BDI model) produces goals that the agent has to fol-

low. A goal could be: to answer a user’s (or peer)

request, to update its own TRat and TRec, etc. These

TRUST MODEL FOR HIGH QUALITY RECOMMENDATION

525

goals are also inﬂuenced by the “User Proﬁle” (this

includes the user’s context). In order to fulﬁl these

Desires (or goals), the delegate agent has to formulate

a set of Intentions, which will become actions. These

Intentions are dictated and later executed by the “Rec-

ommender” and the “Argumentation Engine”. As a

result of these actions, the knowledge (here, TRat and

TRec) of the delegate agent is updated, which con-

stitutes the Beliefs of the agent. Based on these new

Beliefs, more Intentions have to be processed (if the

current goal is not yet satisﬁed) or a new goal is set

(or updated). The same cycle continues as long as

there are goals to be achieved. More details about this

architecture are available in (Sahli et al., 2008).

Goals

Generator

Memory

Communication

Component

User Profile Recommender

Argumentation

Engine

Beliefs

Generator

Beliefs (TVC

graph)

Info

Feedback/info

Info that needs

justification

Argument

Update

(learned)

Profile update

Initial goal

TVC graph

Ask agent/answer agent/ answer user

Goal

Trust rulesTrust Model

Trust rules

TVC graph

Profile

Inform community about user’s feedback

Request from

user/agent

Memory

Goals

Generator

Recommender

User Profile

Beliefs

(TRec/TRat)

Argumentation

Engine

Beliefs

Generator

TRec/TRat

Trust rules

Trust Model

updates

profile

TRec/TRat

Initial goal

Goals

User

Feedback

Profile

Request

Answer

Argument

Feedback

profile

[if justification

needed]

Other Delegate Agents in

the Virtual Agora

Figure 4: The delegate agent’s BDI architecture.

Extension to Mobile Users. We now demonstrate

the easy extensibility of our architecture to support

mobile users. Since delegate agents have to re-

quest/argue about opinions to ensure high-quality rec-

ommendations, they obviously require more interac-

tion. However, when users are mobile, exchanging

these messages between peers (mobile users) would

generate a large and costly wireless communication

trafﬁc. It is thus necessary to avoid remote messages

as much as possible and allow most of communica-

tion to be held locally (peers exchanging messages

should be located at the same server). The VA con-

cept seems to be appropriate to fulﬁl this requirement

since it constitutes a meeting infrastructure where all

delegate agents can exchange local messages. But

how to make the link with mobile users? To achieve

this goal, we extend our agent-based architecture by

assigning a second agent (in addition to the delegate

agent) to each user (here mobile user). We call this

agent embedded agent. This light (has few data and

functionalities) agent is a proxy between the user and

the delegate agent. It is embedded in the mobile de-

vice of the user. It mainly (i) notiﬁes delegate agent

about the user’s feedback, tags (e.g., ratings), changes

of interests or preferences, etc., and (ii) requests rec-

ommendations on behalf of the user. While delegate

agent is deliberative, embedded agent is more a reac-

tive agent. Indeed, it does not support any reasoning,

it is only making the bridge between the user and the

delegate agent and reacting to incoming events. The

architecture described previously is thus extended as

shown in Figure 5. More details about the internal ar-

chitecture of this agent are presented in (Sahli et al.,

2008).

Implementation. We have already implemented a

prototype of the VA. We chose JADEX as a devel-

opment environment. Besides the fact that JADEX is

built over the reliable environment Jade, it handles the

BDI concept which is very useful in our case to eas-

ily implement the VA’s members. We are currently

working in integrating the automatic capturing of the

user’s context to the architecture. We are also work-

ing on implementing the argumentation mechanism

as described in this paper.

Figure 5: Simpliﬁed architecture supporting mobile users.

7 RELATED WORK

Very few papers have addressed the use of argumenta-

tion in their trust models. Stranders (Stranders et al.,

2007) seems to be the ﬁrst to use a form of argumenta-

tion for trust evaluation viewed as a decision process,

taking inspiration from (Amgoud and Prade, 2004).

Prade proposed a similar approach in (Prade, 2007).

However, in these systems, argumentation is an in-

ternal process (internal to each agent) used to sup-

port decision making. Each agent evaluates its trust

in a peer (or a source) according to a set of argu-

ments. In contrast, in our approach, each agent uses

argumentation as a negotiation mechanism with other

agents in order to debate about the “strength” of a

given item. To our knowledge, only Bentahar and

SECRYPT 2008 - International Conference on Security and Cryptography

526

Meyer (Bentahar and Meyer, 2007) have addressed an

argumentation-based negotiation to enhance the trust

model of agents. Their approach is, by the way, com-

plementary to ours. First the goal is to provide a se-

cure environment for agent negotiation within multi-

agent systems. Second, the overal trust degree of an

agent is estimated according to a probabilistic model

which depends upon the number of past and current

accepted arguments.

Our work is also related to numerous works that

have been addressing distributed trust management in

the last years. Our solution of making trust dependent

of past experience and recommendations is a com-

mon choice in distributed trust management (e.g., see

the surveys (Ruohomaa and Kutvonen, 2005; Jøsang

et al., 2005)).

Our design is also related to recommender sys-

tems and collaborative ﬁlter solutions. The abstract

functions we use to initiate trust (i.e., estimate,

eval) and to tag items (i.e., tag) ﬁnd their concrete

counterpart in the solutions that exist to estimate the

trustworthiness of objects during the bootstrap phase

of trust. Again a pletora of research works is avail-

able in the literature. Quercia et al. propose to

use a user’s local information (on the ratings of its

past experiences) to estimate the trustworthiness of

an item (Quercia et al., 2007b). In (Quercia et al.,

2007a), they also design a distributed algorithm that

users can run to predict their referral trust in content

producers from whom they have never received con-

tent before. Both the solutions are suitable for mobile

devices, which make them attractive in our design.

Argumentation has been proposed to enhance

critic systems and recommender systems (Chesnevar

et al., 2006). Potential suggestions which follow a

user’s web query, are ordered by relevance. The rel-

evance relation is deﬁned in terms of argumentation,

whose rules express the user preferences. Warranted

suggestions come ﬁrst, then undecided followed by

defeated suggestions. Differently from our approach,

Defeasible Logic is used to express a user’s prefer-

ences, which can be quite naturally expressed as a de-

feasible reasoning. In this case, argumentation is not

a dialogue between two subjects (as in our approach)

but it is the result of attacks/defenses pairs that arise

in the subjective defeasible reasoning of a user when

she/he evaluate a piece of information retrieved from

the web. Argumentation has been also proposed in

agent systems to reason about desires of agents (Rot-

stein et al., 2007). Here, Defeasible Logic is used to

select, among a set of potentially conﬂicting agent’s

desires, an appropriate one that ﬁts the particular sit-

uation of a given agent.

8 CONCLUSIONS AND FUTURE

WORK

In this paper, we enhance the decentralised recom-

mender system we proposed in (Lenzini et al., 2008)

by integrating argumentation into the trust model.

Within a virtual community of raters, a member can

thus select recommendations not only based on its

trust in these peers, but also on the justiﬁcations they

give. We thus make a member take advantage of rec-

ommendations coming from less trustworthy mem-

bers (e.g., those with which she/he has no direct ex-

perience) if they are strongly justiﬁed by convincing

arguments. As future work, we intend to propose a

full implementation of the argumentation model. This

implies ﬁnding a way to capture the user’s way of rea-

soning (in a user-friendly manner) and integrating the

output in the formal model. An eventual commercial-

isation of our recommender system will thus depend

on linking the formal model with the real data (cap-

tured from the user) via a user-friendly interface.

With respect to the supporting architecture, our

decentralised recommender system is designed based

on the agent paradigm. In fact, we model raters

by autonomous and unobtrusive agents, which frees

users from managing their trust in other members.

Moreover, if users are mobile, we decouple each

agent into two agents (delegate and embedded),

which dramatically reduces wireless trafﬁc between

peers. We are currently working on integrating Con-

textWatcher (Koolwaaij et al., 2006) into our system

in order to effectively support the context. The Con-

textWatcher is a mobile application developed in our

research laboratory, and which aims at making it easy

for an end-user to automatically record, store, and use

context information. It automatically captures some

context aspects such as time, location, and social en-

vironment (of the end-user).

REFERENCES

Amgoud, L. and Prade, H. (2004). Using arguments for

making decisions: a possibilistic logic approach. In

Proc. of the 20th conference on Uncertainty in artiﬁ-

cial intelligence AUAI’04, Banff, Canada, pages 10–

17. AUAI Press.

Barber, K. S. and Kim, J. (2002). Soft security: Isolat-

ing unreliable agents from society. In Proc. of the

5th Workshop on Deception, Fraud and Trust in Agent

Societies, AAMAS 2002, July 15-19, Bologna, Italy,

pages 8–17. ACM.

Bentahar, J. and Meyer, J. J. C. (2007). A new quantita-

tive trust model for negotiating agents using argumen-

tation. Int. Journal of Computer Science & Applica-

tions, 4(2):39–50.

TRUST MODEL FOR HIGH QUALITY RECOMMENDATION

527

Chesnevar, C. I., Maguitman, A. G., and Simari, G. R.

(2006). Argument-based critics and recommenders: a

qualitative perspective on user support systems. Data

Knowl. Eng., 59(2):293–319.

Garcia, A. and Simari, G. R. (2002). Defeasible logic pro-

gramming: An argumentative approach. Theory and

Practice of Logic Programming, 4(1):95–138.

Jøsang, A., Gray, L., and Kinateder, M. (2006). Simpliﬁca-

tion and analysis of transitive trust networks. Web In-

telligence and Agent Systems Journal, 4(2):139–161.

Jøsang, A., Ismail, R., and Boyd, C. (2005). A survey of

trust and reputation systems for online service provi-

sion. Decision Support Systems. (available on line on

ScienceDirect) in press.

Koolwaaij, J., Tarlano, A., Luther, M., Nurmi, P., Mrohs, B.,

Battestini, A., and Vaidya, R. (2006). Contextwatcher

- sharing context information in everyday life. In Proc.

of the IASTED Int. Conf. on Web Technologies, Ap-

plications, and Services (WTAS2006), Calgary, CA,

pages 39–60. ACTA Press.

Lenzini, G., Sahli, N., and Eertink, H. (2008). Agents se-

lecting trustworthy recommenders in mobile virtual

communities. In Proc. of the 11th Int. Workshop on

Trust in Agent Societies (TRUST08), AAMAS 2008,

pages 94–104.

Miller, B., Konstan, J., and Riedl, J. (2004). Toward a per-

sonal recommender system. ACM Transactions on In-

formation Systems, 22(3):437–476.

Nute, D. (1994). Defeasible logic. In Gabbay, D., Hogger,

C. J., and Robinson, J. A., editors, Handbook of Logic

in Artiﬁcial Intelligence and Logic Programming, Vol-

ume 3: Nonmonotonic Reasoning and Uncertain Rea-

soning, pages 353–395. Oxford University Press, Ox-

ford.

Prade, H. (2007). A qualitative bipolar argumentative view

of trust. In Prade, H. and Subrahmanian, V. S., editors,

Proc. of the Int. Conf. on Scalable Uncertainty Man-

agement (SUM), Washington, 10-12 Oct 2007, DC,

USA, volume 4472 of LNAI, pages 268–276. Springer-

Verlag.

Prakken, H. (2006). Formal systems for persuasion dia-

logue. Knowledge Engineering Review, 21(2):163–

188.

Preece, J. (2000). Online Communities: Designing Usabil-

ity, Supporting Sociability. John Wiley and Sons.

Quercia, D., Hailes, S., and Capra, L. (2007a). Lightweight

distributed trust propagation. In Proc. of the 7

IEEE

International Conference on Data Mining, 28-31 Oct

2007, Omaha, NE, USA, pages 282–291. IEEE Com-

puter Society.

Quercia, D., Hailes, S., and Capra, L. (2007b). TRULLO

- local trust bootstrapping for ubiquitous devices. In

Proc. of the 4

IEEE International Conference on

Mobile and Ubiquitous Systems: Computing, Net-

working and Services (Mobiquitous), Philadelphia,

US.

Rao, A. S. and Georgeff, M. (1995). Bdi agents : from

theory to practice. In Proc. of the 1st Int. Conference

on Multi-Agent Systems (ICMAS95), June 12-14, San

Francisco, USA, pages 312–319. AAAI Press.

Rotstein, N. D., Garc

ıa, A. J., and Simari, G. (2007). Rea-

soning from desires to intentions, a dialectical frame-

work. In Proc. of the 22nd. AAAI Conference on Ar-

tiﬁcial Intelligence (AAAI 2007), Vancouver, Canada,

22-26 July 2007, pages 136–141. AAAI Press.

Ruohomaa, S. and Kutvonen, L. (2005). Trust manage-

ment survey. In Proc. of the iTrust 3rd Int. Conference

on Trust Management, 23-26, May, 2005, Rocquen-

court, France, volume 3477 of LNCS, pages 77–92.

Springer-Verlag.

Sabater, J. and Sierra, C. (2002). Social regret, a reputation

model based on social relations. SIGecom Exchanges,

3(1):44–56.

Sahli, N., Lenzini, G., and Eertink, H. (2008). Trustwor-

thy agent-based recommender system in a mobile p2p

environment. In Proc. of the 7th Int. Workshop on

Agents and Peer-to-Peer Computing (AP2PC08), AA-

MAS 2008, pages 1–11.

Stranders, R., de Weerdt, M., and Witteveen, C. (2007).

Fuzzy argumentation for trust. In Sadri, F. and Satoh,

K., editors, Proc. of the 8th Work. on Computational

Logic in Multi-Agent Systems (CLIMA VIII), 10-11

Sep 2007, Porto, Portugal, LNAI. Springer-Verlag.

(in press).

Teacy, W., Patel, J., Jennings, N., and Luck, M. (2005).

Coping with inaccurate reputation sources: Experi-

mental analysis of a probabilistic trust model. In Proc.

of the 4th Int. Joint Conf. AAMAS, pages 997–1004.

Toivonen, S., Lenzini, G., and Uusitalo, I. (2006). Context-

aware trustworthiness evaluation with indirect knowl-

edge. In Proc. of the 2nd Semantics Web Policy Work-

shop (SWPW 2006), CEUR Workshop Proceedings.

CEUR-WS.org.

SECRYPT 2008 - International Conference on Security and Cryptography

528