Commonly, when a Trustor evaluates a Trustee's trustworthiness, it is assumed that the evaluation is based on information directly available to the Trustor. This can concern for example the reputation and recommendations characterizing the Trustee. In cases of context-aware trust, this is further restricted by concentrating mainly on information in a similar enough context as is effective at trust evaluation time. However, this information is not necessarily available to the Trustor. Surprisingly, in such scenarios the literature suggests either to wait for someone else to collect the needed experience, or to trust blindly. In this paper, we discuss solutions that help the Trustor to conduct its evaluation even if direct knowledge about the Trustee is lacking. We approach this by allowing the Trustor to make use of networks connecting the Trustor and the Trustee, as well as the context information characterizing the entities appearing in these networks.
Trust is an increasingly important phenomenon to grasp and support in open
environments, such as the Internet, where participants are not necessarily in direct contact with
each other. A common scenario is that a the subject of trust (Trustor) is searching for
a service or a product (Trustee) for a certain purpose. Semi-automatic trustworthiness
evaluation is of special relevance on the Semantic Web, where Trustors can be software
agents in addition to human beings, and Trustees are software agents or web pages
carrying information for Trustors to depend on. To perform an appropriate evaluation,
Trustors request Trustees’ credentials, often expressed in terms of profiles, reputation
descriptions, and recommendations (cf. [
Context-awareness is also an emerging computer science trend, which takes
situational details into account. Generally, in computer science context refers to any
information characterizing the situation of any entities considered relevant to the interaction
between a user and an application, including the user and the application themselves,
as well as their surroundings [
Many research efforts in addition to ours also acknowledge that context information
may help to define trust credentials (cf. [
Trust management frameworks operate under the assumption that the Trustor can
directly access the information he requires to complete the trustworthiness
evaluation [
In this paper, we study context-aware trust establishment by considering scenarios where direct information about the Trustee is not necessarily available to the Trustor. We claim that even in such situations there are better options for Trustors to choose from than to trust/distrust blindly. For example, the Trustor can evaluate the trustworthiness of another entity somehow related to the Trustee. In many real situations humans act like this. We trust a car manufactured in a certain country, if our previous experiences with cars manufactured in that particular country are good, even if we have no experiences of that particular make. In many cases this kind of indirect evaluation suffices to accomplish a fair judgment to start with.
The particular cases we consider are the following : (i) Trustee’s behavior across
contexts is unknown to the Trustor, meaning that the Trustor has no previous
knowledge of any behavior of the Trustee; (ii) Trustee’s behavior in the current context is
unknown to the Trustor, meaning that the Trustor might know the Trustee, but not how
the Trustee behaves in the current context; (iii) Trustee’s recommender and/or
recommendations are unknown or unaccessible to the Trustor. Cases (i) and (ii) are targeted to
reputation information, as they are dependent on the Trustor’s knowledge and opinions
on past states-of-affairs. Case (iii) relies on recommendations available to the Trustor,
although the mechanisms to be considered in terms of (i) and (ii) could be plugged in it
too. Note that we consider the context to be fully observable [
The rest of the paper is organized as follows. In Section 2, we present some relevant related work. In order to pinpoint the contribution of this paper, in Section 3 we then present the baseline case where there is complete and direct information influencing trustworthiness evaluation available. We also formalize operational semantics for the trustworthiness evaluation process; it will help us later on to discuss the changes in the trustworthiness evaluation process when only indirect information is available. In Section 4, we delve into the scenarios where the Trustor has little or no reputation knowledge about the Trustee. Section 5 considers the case where recommender is not known to the Trustor. Finally, Section 6 concludes the paper and outlines some future work. 2
The interaction between trust and context has attracted the attention of researchers
only recently, and from different perspectives. In the Web Services domain, for example,
context is used to anonymize the authentication procedure [
In [
The problem of inferring trust from recommendations has appeared in the
literature for a long time. Yahalom et al. [
Richardson et al. explicitly address belief composition in the Semantic Web
domain [
O’Hara et al. analyze costs and benefits in different paradigms (optimism,
pessimism, centralized, investigation, and transitivity) of dealing with trust in Semantic
Web [
Baseline: Direct Information Available to the Trustor
This section summarizes the formal definitions of context-aware trust evaluation
functions we introduced in [
From Context-independent to Context-aware Trust Evaluation Functions
In [
Function (1) inputs a description of the Trustee in terms of the following parameters: (a) a set Q ∈ Quality of Trustee’s quality attributes; (b) a trust value m ∈ TValues; (c) a set M ⊆ TValues of trust values. Set Q models any information that A knows directly about Trustee, such as the Trustee’s profile. Value m models the Trustee’s reputation in the viewpoint of A, that is a trust value stored in A’s local space. Set M represents recommendations, which are Trustee’s trust values based on the viewpoints of recommenders.
It is recognized that trust changes over time [
trust0A,σ(B) := trustA,σ(Q0B, ⊥, MB0 ) trustiA,σ(B) := trustA,σ(QiB, miB, M Bi) wmheaerseurQe,iBm∈iB Q∈uaTlViatlyueasreisthtehequreapliutytaatitotrniboufteBs (orfecBo matmtiemnedait,io⊥niisnaAn’usnvdieewfinpeodinttr)uastt time i, and M Bi ⊆ TValues are recommendations on B at time i.
Definitions (1) and (2) can be extended to deal with context. Their context-aware
counterpart is written as follows [
C
rt;(i,m)
A −−−−−→ B
Martinelli [
C
it;(i,m) A◦−−−−−→ B
R,C Here, R is the set of recommenders whose opinion has been considered when composing m. The semantics of context-aware trust evaluation is defined as an inference system, as depicted in Figure 1. We now comment each rule separately.
Rule (5) defines the scheme of our inference system’s axioms. If A’s subjective evaluation of B’s qualities at time i evaluates to m and if C is the context available at time i, then A trusts B in measure m′ = C ⊙ m, where the operator ⊙ is that of equation (2′). Premises in brackets (e.g., [trustA(QiB )] = m) are evaluated at a meta level.
Rules (6) formalize the operational management of recommendations. In particular,
rule (6.a) shows that an indirect trust on B derives from A’s referral trust in D and
from the (direct) trust that D already has in B; rule (6.b) and rule (6.c) show how
to concatenate referral trust along a chain of reference and how to aggregate indirect
trust across multiple paths of recommendations, respectively. Accordingly to [
Note 1. The semantics of rules (6) are incomplete unless we give the semantics of the two operators ⊗ and ⊕.
Reasonably, ⊕ must be at least associative and commutative (to be order-independent)
and ⊗ at least associative (along a chain of recommendations). Some authors (e.g.,
[
Note 2. Relation ≡ ⊆ Context × Context needs to be instantiated to complete the semantics of rules (6) and (7).
In its simplest form, ≡ interprets as identity: a reputation or recommendation is adequate only if performed in the same context. Alternatively, ≡ can be an equivalence relation between contexts—only experience performed within an equivalent context can contribute to present trust—or ≡ can be a reflexive and symmetric relation modeling a semantic closeness. For example, if d is a distance between contexts, ≡ can be d(C, C′) ≤ r, where r is the radius of the neighborhood. In case ≡ is not the identity, it is reasonable to expect the derived trust to be < m ⊕ m′. Closer study of this modified version of the rule is left as future work.
Rules (7) define how to obtain direct functional trust. More specifically, rule (7.a) models the aggregation of a direct functional trust. Rule (7.b) models our approach of dealing with reputation as a (direct) past experience that is combined with the present direct trust. Similarly to the recommendation rules, here context acts as a filter in favor of those experiences occurred in a ≡-related context. Finally, rule (7.c) states that a past experience can be used as if it was a new experience presently, at the price of some trust decay (here represented by the constant discount ω).
Note 3. In rules (7.b) and (7.c) constraints over time can guide the search strategy in the past. Each strategy reflects a different attitude in considering reputation (e.g., choosing a maximal j implies the consideration of most recent experience stored in the reputation base).
Rules (8) define functional trust (the goal of our proof system) as a generalization of direct and indirect trust.
As a final remark, we observe that our inference systems allows different proof searches with different result for the same goal. Various implementations and optimization strategies are possible, but we do not discuss them in this paper. 4
So far we have implicitly assumed that the Trustee’s quality and contextual attributes needed in order to evaluate trust are directly available to the Trustor. In real situations, we may be obliged to relax this assumption. Consider, for example, a situation where we would like to evaluate the quality of a new scientific conference. Due to its newness, the conference is not ranked yet. Moreover, we will not find anyone known to us recommending it either. In such a situation, we basically have only two alternatives: to give up the evaluation (i.e., blindly trust/distrust), or to look for and rely on indirect information. For example, we can evaluate the prestige of the publisher of the conference (RECOMMENDATION-RULES) A −r−t−;(−i,−m→) D
D◦−d−t;−(i−−−1−,m−′→) B
C A◦−i−t;−(i−,m−−⊗−m−→′) B {D},C i > 0 (b)
A −r−t−;(−i,−m→) D
D −r−t;−(i−,m−→′) B A −r−t−;(−i,−m−⊗−m−→′) B A◦−i−t;−(i−,m−→) B A◦−i−t;−(i−,m−→′) B
R,C R′,C
A◦−i−t;−(i−,m−−⊕−m−′→) B
R∪R′,C (d)
A◦−d−t−;(−i,−m→) B
C
A◦−i−t;−(i−,m−′→) B [C′ ≡ C]
R,C′ A◦−d−t−;(−i,−m−⊕−m−→′) B
C (a)
A◦−d−t−;(−i,−m→) B A◦−d−t;−(i−,m−→′) B
C C A◦−d−t−;(−i,−m−⊕−m−→′) B
C (b)
A◦−d−t−;(−i,−m→) B
C
A◦−d−t−;(−j,−m−→′) B [C′ ≡ C]
C′ A◦−d−t−;(−i,−m−⊕−m−′→) B
C (ADDITIONAL-RULES) (a)
A◦−d−t−;(−i,−m→) B
C A◦−(−i,−m→) B
C j < i
(c) (b)
A◦−i−t;−(i−,m−→) B
R,C A◦−(−i,−m→) B
C
A◦−d−t−;(−i−−1−,m−→) B
C A◦−d−t−;(−i,ω−⊗−−m→) B
C i > 0 (5) (6) (7) (8) proceedings, or we can look for the reputation of its program chairs and committees. In the case of a new workshop colocated with a conference having a history, we can also consider the quality of the conference when evaluating the workshop. This section studies how trust can be evaluated in such situations. If we come across a Trustee not known to us, that is, we possess no prior reputation information about the Trustee, how should we go about evaluating the trustworthiness? One well-known solution in the literature is to ask for recommendations. In Section 5 we discuss recommendations and how to deal with them. Here, instead, we analyze a complementary solution, namely utilizing direct information of entities known to the Trustor and “related” to the Trustee (see Figure 2 (a)). Let us consider again the example about evaluating the trustworthiness of a new scientific conference. Due to the absence of any information about the conference, we can find it satisfactory to evaluate the trustworthiness of the conference proceedings publisher, as well as those of the program chairs and committee members.
From a formal point of view, the previous solution is expressed by the following additional (to the INIT-RULE) inference rule (9) where a trust relationship with B in a certain context C is deduced by a trust relationship with another Trustee “related” to B in the same context.
dt;(i,m) A◦−−−−−→ D
C
[D ∼ B]
Here, the semantics of the rule requires us to instantiate the relation ∼; it can be an
equivalence relation, or a reflexive and symmetric relation among Trustees that defines
the concept of entity neighborhood. For example, Figure 2 (a) suggests the use of a
measure of closeness among entities (see also Section 5). In this case A ∼ B if and only
if cls(A, B) ≥ th where th is a threshold. In the following, we assume the closeness
metric ranging in [
C trusworthiness of the Trustee (Figure 2 (a)). However, it is envisaged that often more appropriate results can be obtained by considering the Trustee itself, and its behavior in contexts which are similar enough with the current one (Figure 2 (b)).
Let us continue with the scientific conference example, but this time from the conference chair’s point of view. Suppose that the chair is gathering a program committee for the new conference. Here subject areas of the conference call for papers constitute the relevant attributes, which guide the conference chair in inviting appropriate members for the program committee. More specifically, the chair has two major options: In the case of previous conference chair experience in similar enough conferences, the chair can go about evaluating the performance of the PC members in those conferences and make up his mind based on that. Alternatively, the chair can look up other good and similar enough conferences, and count the most frequent PC members and invite them to join.
From a formal point of view, the previous solution is expressed by the following additional rule (as part of REPUTATION-RULES):
dt;(i,m) A◦−−−−−→ B
C′
[C′ ≡ C] Again, ≡ can be an equivalence relation, or a reflexive and symmetric relation among contexts that defines the concept of context neighborhood. Figure 2(b) suggests one implementation of relation ≡ based on context similarity; C ≡ C′ when cls(C, C′) is greater than a threshold th. The inferred trust value m′, here left unspecified, reasonably depends on m and on the nature of the relationship between C′ and C. For example, m′ = ωs ⊗ m where ωs = (s, 1 − s, 0) is the discount build from s = cls(C, C′) between C and C′. Here, m′ can be computed either as ⊕kmk (e.g., the consensus among all the trust values) or as the trust value of the context that has maximal similarity with C.
In rules (6) recommendations carry the context C′ they relate to. Recommendations are considered only if ≡-related with the current context C. Dealing with recommendations in this way is possible only if the Trustor knows the recommenders. We now loosen this requirement. In essence, we allow entities not directly known to the Trustor to be included in the trustworthiness evaluation process as recommenders. In this case, a Trustor may deduce indirect trust directly from an entity, if the entity is “close enough” to the Trustor. In other words, referral trust is approximated by the semantic distance between entities, with the intuitive meaning that “the closer, the more trusted”. Formally, this new evaluation step is synthesized by the following variant of rule (6.a): (10′)
dt;(i−1,m) D◦−−−−−−−→ B
C it;(i,m′) A◦−−−−−→ B {D},C Here, the calculus of m′ depends on the nature of the relation between A and D; for example, m′ = ωs ⊗ m where ωs is the discount (s, 1 − s, 0) that reflects the closeness s between A and D. The relative importance of a given recommender is estimated based on its relation with the Trustor.
The closeness between two entities can be grounded on the number of links between the Trustor and the recommender. Figure 3 depicts this. Note that there can be multiple parallel paths from the Trustor to the recommender, and they can be taken into account in differing ways. Only the shortest path can be considered, or alternatively all (or some reasonable amount of the) paths can be included in the calculation. The underlying idea is that the more paths there are between the Trustor and the recommender and the shorter they are, the more relevant the recommender is in the eye of the Trustor. Closeness is expressed by the following formula: cls(A, D) = X k∈I
1 trustee
recommender recommends trustor avg(rec x L x E) links L={L1,L2,...,Ln}} where p1, p2, ..., pn ∈ P is the ordered set of alternative parallel paths found between the Trustor and the recommender so that |p1| indicates the number of links in the shortest path, |p2| in the second-shortest, and so on.
Note that there can be multiple paths that have the same amount of links. As a representative for each set of paths that have an equal amount of links we choose the path with the smallest index. An ordered set of indexes I is created so that only the indexes of the representatives ∈ I. If all paths ∈ P have a different amount of links, then I = {1, . . . , n}. With ♯|pk| (for all k ∈ I) we mark the number of paths, in the set of equal length paths, represented by pk.
As an illustrative example, consider again the conference chair as Trustor A and the proposed PC member’s colleague or boss as recommender D and two paths between them. One of the paths has one link and the other two. (If only the shortest path was considered, the closeness metric of the recommender would be 211 = .5). The closeness metric taking into account both paths is 211 + 312 = 23 ≈ .6·7, and I = {1, 2}. If we add yet another path to the picture, this· time ·with five links, the closeness metric is 21·1 + 31·2 + 61·3 = 1183 ≈ .72. Here I = {1, 2, 3}. Now, consider there are three paths between the Trustor A and Trustee D, two having one link each and one having five links. The set of indexes becomes I = {1, 3} and the closeness metric becomes √212·1 + 61·3 ≈ .76
Two main approaches concerning different link kinds can be distinguished. In the first of these approaches, all link kinds L1, L2, ..., Ln ∈ L—be they based on profession, kin, plain acquaintance, and so on—are considered as equally important with regard to the trustworthiness evaluation. The second, in turn, makes distinctions between different link kinds and values some over others. For example, with regard to the program committee membership, professional links can be put more emphasis than acquaintanceships or family relations.
To make distinctions between different link kinds ∈ L we add a weighting to them. Let wpkj ∈ R be a weighting for a link in path pk, where j = 1, . . . , |pk|. The mean link weight for path pk is defined as
Wpk =
P|jp=ki| wpkj .
|pk|
For a set of paths P = p1, p2, . . . , pn, we normalize the path weights Wpk to [
Wp′k =
Wpk max{Wpj | j = 1, . . . , n} .
In case there are paths that have an equal amount of links, the mean of their normalized path weights is used. Finally, the weighted closeness metric wcls(A, D) becomes wcls(A, D) = X k∈I
Wp′k
Let us continue with the conference example. Suppose we have the same two paths between the Trustor A and recommender D as earlier. But now the shorter path consists of one link of type “family relation”, weighted at 2.5, whereas the path with two links consists of professional links with corresponding weights 4 and 6. The mean link weight for the shorter path is 2.5, and 5 for the longer path. The normalized link weights are 1 thus 12 and 1, respectively. The weighted closeness metric of these paths is 221 + 312 = 152 ≈ .42. Suppose that at a later time the family member whose relation was· weig·hted at 2.5 becomes an assistant, and the weight of this relation is 4. In this case the weighted 4 distance metric of these paths becomes 251 + 312 = 1370 ≈ .57.
· · 6
We described and formalized means for evaluating trustworthiness in cases where the Trustor does not possess direct information about the Trustee. We considered both the absence of direct reputation information, that is, lack of Trustor’s personal experiences of the Trustee, and the absence of direct recommendation information, that is, lack of recommendations transmitted to the Trustor by entities known to the Trustor. We discussed cases where the Trustee/Recommender is unknown to the Trustor across contexts, meaning that the Trustor has no knowledge whatsoever about the actions taken by the Trustee/Recommender. In addition, we considered cases where the Trustor has some knowledge about the Trustee/Recommender, but not in the current context.
As a solution we propose to use measures of similarities among entities, and among contexts. Similar entities to the Trustee and a recommender can be used instead, in case Trustee and/or recommenders are unreachable to the Trustor. Additionally, the Trustor can search for a Trustee’s reputation in a similar context, if information concerning the Trustee’s reputation in the present context is missing. Whilst formalizing our approach, we illustrated its usage via a running example.
Our future work around the area includes further investigating the relationships between the Trustor and the Trustee. Research questions are for example comparing different similarity metrics connecting the Trustor with the Trustee (via multiple paths containing recommenders and other acquaintances, as well as varying contexts). In addition, we plan to empirically test and evaluate these metrics.
This work has been supported by the EU-ITEA project “Trust4All”. In addition to the project consortium and financiers, the authors would like to thank the anonymous reviewers of the 2nd International Semantic Web Policy Workshop (SWPW’06), as well as Sami Kauppinen for his useful comments on the final version of the paper.