The Impact of Competence and Benevolence in a Computational Model of Trust

. Trust is a fundamental element of any social network. However, despite numerous studies on trust, few have conducted studies across disciplines to provide a complete picture of the diﬀerent dimensions of trustworthiness, such as integrity, competence and benevolence. In this paper, we focus on two of these dimensions, competence and benevolence. We propose techniques to evaluate the competence of the trustee in speciﬁc situations and infer the benevolence of the trustor towards the trustee when the trust evaluation is made. Moreover, we evaluate both competence and benevolence on the perceived trustwor-thiness of the trustee, taking into consideration the development of the relationship between the trustor and the trustee over time. We identi-ﬁed diﬀerent stages in this relationship development and use them to evaluate trustworthiness of trustee in the absence of evidence that could otherwise be used to evaluate trustworthiness. Finally, we set an experimental scenario implemented as an agent-based model to evaluate our approach. The results obtained from these experiments show that the proposed techniques can improve the reliability of the estimation of the trustworthiness of the agents.


Introduction
Trust is considered a fundamental basis in social societies.The social networks, however, evolve over time therefore, we need a sophisticated method that enables an agent to select the trusted peer to interact with.Trust has both a backward looking as well as a forward looking dimension, i.e. one expects certain behavior in the future based upon pastime experiences.Since one can never be sure if that the conditions in the past are exactly the same as they are now (e.g, intentions of agents may change over time one may have to trust someone that had no experience with etc.), the estimation of trustworthiness is predominant to assess the trust between each peer in such network.An agent-based model (ABM) of such societies requires a computational model of trust.Trust is a concept that will always need contextualization.It is something completely different to trust that someone will pay his debts, not commit a crime or is able to solve a complex mathematical calculation.In our pursuit for a computational model of trust that can be used in ABMs to allow us to study (non-)compliance in societies we focus on the trustworthiness of an individual in performing a certain task in a given situation, taking into account the agent's competence in the matter, his overall integrity, and the stage of his relationships with the trustor.In order to better estimate the trustworthiness of trustees, it is important to consider these three dimensions individually, and to combine them in a dynamic way by taking into consideration the situation and the development of the relationship.However, the majority of the computational trust approaches presented in literature [1] estimates the trustworthiness of agents as a block and does not distinguish between these trustees' factors.In this paper, we present a computational trust approach grounded on the trust model that presented by Marsh [2].We will show that the resulting computational trust model is able to capture the competence and benevolence of the trustee.To estimate the trustee's benevolence and competence, we use the evidence available to the trustee by considering different situations.Our approach combines the estimated competence of the trustee with the estimated benevolence into one single trustworthiness score.This score reflects the development of the relationship between trustor and trustee at the time of the assessment take into consideration.To estimate the competence of the trustee, we introduce three different stages of relationships between trustor and trustee.The results we obtain and present in this paper are very encouraging, as they show that there is a clear benefit in considering different stage of relationships between trustee and trustor in the described situations: the benevolence enhanced trust models allowed for a more accurate estimation of the trustees' trustworthiness than the original computational trust models The remainder of the paper is organized as follows.Section 2 highlights the trust model and its antecedents.Then, Section 4 introduces our social computational model with its components.Section 3 deals with the simulation and results of the proposed solutions.Related work is covered by Section 5. Finally, Section 6 concludes the paper.

Trust
Trust is an essential part of social interaction.Trust is a broad concept studied in areas such as sociology and psychology [3].The concept of trust has received ample attention from various disciplines, and although prior research has put forth diverse interpretations of trust, a common core emerges4 Following Mayer et al. [3] we use the following trust definition: "Trust is the willingness of a party to be vulnerable to the actions of another party based on the expectation that the other will perform a particular action important to the trustor, irrespective of the ability to monitor or control other party [3]".Building on this definition, we define inter-organizational trust as the expectation held by one firm that another will not exploit its vulnerabilities when faced with the opportunity to do so [3,5,6].This expectation is confirmed when parties: demonstrate competence relate to the potential ability of the evaluated entity to do a given task, act accordingly to fulfill the commitments even when acting on them is not in self interest and accept the consequences, and do good and act out of kindness even if unforeseen contingencies arise.
Our definition thus bases inter-organizational trust on three related components: competence, integrity, and benevolence, which have been proposed by Mayer et al. [3].Computational trust is considered as an enabler of technology in virtual societies, and the estimation of trustworthiness is paramount to assess the trust that a trustor agent has on a given trustee.An individual is more or less trustworthy in performing a task in a given situation depending on his ability in the matter, his overall integrity, and the stage of his relationships with the trustor.Therefore, in order to better estimate the trustworthiness of agents, it is important to consider these three dimensions individually, and to combine them in a dynamic way, by taking into consideration the situation and the development of the relationship.However, the majority of the computational trust approaches presented in literature estimate the trustworthiness of agents as a block and do not distinguish between these trustees' attributions.In following, we present a computational trust approach grounded on multidisciplinary literature on trust that is able to capture the competence and benevolence of the agent under evaluation.This framework has been implemented with ABM.

Computational trust model
In this section, we introduce the social computational trust model.Our aim is to define a mechanism for estimating trustworthiness of a trustee that can be used by the trustor to evaluate trust and make decisions about the future relationship with the trustee.Extracting trustworthiness of trustee based on Mayer et al.'s model [3] has been only implemented by few scholars such as [7].Most of these computational trust approaches estimate the trustees' trustworthiness using individual items of evidence about these trustees' behavior in the past interactions, either with the trustor or with third party agents [8][9][10].However, none of these approaches is able to estimate the benevolence of the trustor towards the trustee.We claim that understanding the benevolence and competence of the trustor towards the trustee at the moment of the trust decision is fundamental for being able to accurately estimate the latter's trustworthiness.With this in mind, we present the main hypothesis of this work as follows; Hypothesis.In long-term and close relationships, effective commitment arises which has a positive impact on the benevolence and competence of partners to perform the given tasks.

Basic Notation
Our generic computational trust model is applied to environments where trustor agents select the best trustees to interact with, having the posterior establishment of dyadic agreements between partners.We define the society of agents by A, where x and y represent as trustor and trustee respectively and x, y ∈ A. We represent T (x,y) as the amount of trust x has upon y based on the realization of a given task t ∈ T with respect to the situation s i , where T = {t 1 , t 2 , ..., t m } is the set of all possible m tasks in the society and s i = {s 1 , s 2 , ..., s i } is the set of all possible situations in the society.These set of tasks bring duties for the trustee that need to be fulfilled.As we mentioned before, context plays an important role in our model.In order to characterize and describe the situation leading to an agreement, we consider the definition of context as including four main types of context: identity, time, location, and activity [11].
where F D (fulfill duty) means the trustor considers that the trustee performed whatever matter he had to perform on time, F DD (fulfill duty with delay) means the trustee was presented with an unexpected delay in the performance of the task (or duty), and V (violation) means the trustee did not perform the given task.Considering O = {F D, F DD, V }, possible values for this function are val(F DD) = 1.0, val(F D) = 0.5, and val(V ) = 0.5.Finally, the set of all the existed evidence on given trustee is represented by where ε represents all evidence available on trustee from the all the trustee's direct neighbor.Following, E (x,y) shows all the evidence about the direct interactions between trustor and trustee

Our social computational trust Model
The benevolence-competence based computational model of trust that we present in this paper is a part of a larger framework of social trust based on Mayer's trust model.It integrates three distinct functions: the competence evaluation function (Com (x,y) : ), the benevolence evaluation function (Ben (x,y) : E (x,y) ∈ [0, 1]), and the trustworthiness evaluation function (T W (x,y) ∈ [0, 1]) in the set of situations si.We describe each of these functions in the following subsections.We illustrated the computational model is illustrated in Figure 2.

Benevolence function
Benevolence is considered as a key element of trust and an antecedent of trustworthiness by several scholars (e.g.[12,13]).The estimated value of the benev-olence of trustor x toward trustee y, Ben (x,y) , is derived from the direct interactions (i.e.E (x,y) ) between trustee and trustor in the situations s i .The output of the benevolence evaluation function Ben (x,y) , defined in [0, 1], is Where S is the set of situations, in which x has interactions with y.

Competence function
The competence evaluation function Com (x,y) estimates the general ability of the trustee under evaluation in performing a given task t in a specific situation s.This function takes all the evidence available on the trustee under evaluation Ex ( * ,y) as input.The output of the competence evaluation function is the estimated competence of the agent, Com (x,y) , defined in [0, 1].Competence, as risk, involves an agent making a judgment about the trustee's ability to perform the given task.We consider three different possible situations to evaluate trustee's ability.
1.There is no evidence available from the trustee.To judge the trustee's competence, the trustor will calculate the risk of trusting a stranger and decide based on the risk as where P r is the probability of performing the task by the given trustee.2. Situation β: there are some evidence but not for the considered context.In this situation, the trustor collects all the evidence from other agents and evaluates the competence of trustee based on them as where T x (y, β) denotes the basic trust that x has on y and β is the set of all situations in which x has interactions with y.The basic trust y calculated as 1 / |N | β∈N T (x,y) .N denotes as the set of situations similar to the present situation (s i ) in which x has interactions with y.
3. Situation α: there is related evidence about the agent in this context or similar situation.
where α is the set of all situations in which x has interactions with y.These three situations (no evidence available, β and α) are assumed to help an individual to make the decision.In this paper, we assume that trustor and trustee collaborate in the similar context (i.e.situation α).

Function T W (x,y)
The trustworthiness evaluation function T W (x,y) takes into consideration the perception of the competence and benevolence of the trustee (Algorithm 1).
In the algorithm 1.We assume that there are at least, two past interactions Algorithm 1 Calculate T W (x,y) Require: E ( * ,y) : the set of all evidence about trustee y Ensure: Ben (x,y) ← F (Ben (x,y) ) 3: Com (x,y) ← F (Com (x,y) ) 4: T W (x,y) = (Com (x,y) + Ben (x,y) ) 5: end if 6: return T W (x,y) between the trustor and the trustee under evaluation.T W (x,y) is the average of the Benevolence and competence of the trustor and the trustee.

Simulation setup
In this set of experiments, we want to test Hypothesis 1, which we reformulated as follows: in long-term and close relationships, effective commitment arises which has a positive impact on the benevolence and competence of partners to perform the given tasks.The experiments were conducted in Jadex [14] environment.We set up a social network, shown in Figure 3 to represent our network.We setup a collaborative network of organizations.This social network represents a collaborative network of organizations like the ones we study in our SARNET research project where service providers collaborate and act on behalf of partners, acts that may harm the individual interests, all in order to protect the collaborative network against cyber attacks.Each node represents an autonomous organization that needs to trust other parties and share sensitive information with them.For the simplicity, we assumed that there is only one task being negotiated by all nodes which mitigates an attack and shares the attack information with other parties 6 .We define four different situations: s 1 provides a specific number of samples within 24 hours, s 2 provides a specific type of resource (e.g.allocating resources), s 3 blocks a link, and s 4 monitors a specific traffic.This model starts after the establishment of an agreement between the trustor and the selected trustee, thus excluding the selection process itself.It focuses on both types of agents' decision concerning the fulfillment of the established agreement: the trustees may opt to fulfill the agreement (the trustors will report the outcome F DD) or to delay its realization.Accordingly, the trustors may respond to a delay by either retaliating, denouncing the breach (reporting outcome V ) or forgiving the contingency (reporting outcome F D).

Result
Our result consists of two parts.First, we calculate benevolence of agent y by considering all the evidence (i.e direct interactions) that x has on y.Second, evaluate the competence of agent given trustees a, z and y in four mentioned situations from the trustor's x view.Hence, we perform four different types of situations simultaneously, each with six agents.We assume that agents are honest and there is no conflict on the evidence and message are encrypted (the interminable agents cannot manipulate the message).In order to compare all approaches, we measure and average the number of agreements with outcomes F D, F DD and V .We are able to calculate the benevolence and competence of each trustor by equation 1 and 4. To calculate the Ben (x,y) , we extract all the evidence that x has on y.We perform the simulation for 20, 50 and 100 rounds of interactions.We also consider that each agent can freely fulfill duty, fulfill duty with delay and violate the agreement.We have summarized the result in Table 1 (including mean M and standard deviation SD).To evaluate the competence function, we select three agents a, z and y from the set of agents and calculate the competence of these three agents from trustor' x view.x will collect (collecting the evidence is done by sending a query to each agent and asking its opinion) all the available evidence from the neighbour's of a, z and y which are reported in Table 2.The simulation has been perform for four different situations and three different rounds.We verified that the benevolence and competence of trustees are increased as the number of interaction increases.For instance, with only 20 rounds, when the number of interactions between any two partners is not large the benevolence is small.By increasing the number of rounds, the benevolence increases significantly.Indeed, this confirms that the number of interactions is, in fact, impact the benevolence existing between any pair of trustor-trustee.In the case of competence, we also observed the same behavior from the simulation.The competence of agent is influence directly by the number of interactions.As we mentioned before, the simulation has been repeated for 50 and 100 rounds.The benevolence value reaches the maximum value of one by increasing the number of interactions between partners and the partners that are considered to be in a close relationship.Indeed, each trustor can conclude the trustworthiness of trustee in each round and make a decision.Overall, we are able to confirm the truthfulness of Hypothesis 1.

Related work
Many computational trust models have been presented by different scholars, nevertheless, only a few models are actually social computational models.Adalie et al. [15] presented a conceptual model of social trust based on Kelton et al.'s model [16].Kelton et al.'s consider ability, positive intentions, ethics, and predictability as the trustworthiness components.Adalie et al. used a probabilistic approach to implement the model, but by recognizing the limits of such approach in the treatment of the social concepts, their model was not impelemented.Among all the presented computational trust models [17], the only computational approach that includes a comprehensive set of features grounded on the theory of trust is the socio-cognitive model of trust by Castelfranchi and Falcone [18].This model considers that the trustor has a goal that can be achieved by the action of the trustee.In their view, trust is formed by considering the different beliefs that the trustor has about the trustee, either internal (beliefs on competence, disposition, and harmfulness) or external (opportunities and dangers).The values of these beliefs are further modulated by meta-beliefs about the relative strength of each belief.The richness of this model makes it hard to implement in practice.In fact, the current implementation of the model (e.g., [18]) requires extensive manual configuration by domain experts for each trustee and task under assessment and oversimplifies the theoretical model.Moreover, it requires explicit information about the competence and disposition (or similar beliefs) of the agent under evaluation, which may be hard to get in dynamic agent-based environments.Another social trust model was presented by Urbano et al. [7] called situationaware and social computational trust Model (SOLUM).Their computational model consists of two parts: the first part is a general framework of computational trust, which is based on two fundamental characteristics of trust, the trustor's disposition and emotional state.They adapted also Mayer's trustworthiness dimension that includes the ability, integrity, and benevolence to determine trust.For the second part, they proposed a set of distinct techniques to extract information about the individual dimensions of the agent's trustwor-thiness from the set of structured evidence available to the agent.The main difference between our model and Urbano's model is that we consider different stages of relationships for the competence function.We slightly adapted and modified the Marsh [2] competence formulization by considering three different situations for trustor to make a decision about the (future) collaboration with the trustee.Finally, Herzig et al. [19] formalized the model of Castefranchi and Falcone, in multi-modal logic, adding the notions of occurrent trust and dispositional trust (i.e., trust in a general disposition of the trustee to perform a similar task some point in the future).Skopik et al. [20] purposed a semi-distributed information sharing platform where different organizations can share the incidents information with their trusted peers.Skopik et al., proposed a fuzzy method to evaluate trust among members.The major aim of social trust in their model is to personalize online interactions and prioritize collaboration with trustworthy individuals.The author claimed that trusted relations can be defined manually by users, e.g., by declaring "friend-relations" or can be determined automatically through mining of interactions.However, their social model is based on personal experience of each member and suffers from the scalability issue.

Conclusion and Discussion
Computational trust is crucial for decision making regarding possible agents' future joint activities such as the alliances.It heavily relies on the estimation of trustworthiness to assess the trust on particular trustees.To better estimate this trustworthiness, it is important to estimate, besides other relevant features, their competence and benevolence separately, and to combine them taking into consideration the particular situation and relationship.In this paper, we described a part of our trust computational model.We evaluated our approach in a simulated experimental environment.We have proved that the trustworthiness estimation grew with the increasing number of interactions between any trustor-trustee pair.Besides, we went beyond the traditional evaluation of computational trust models (such as Fuzzy logic evaluation) and used a model of agents' behavior where both trustors and trustees evolve their behaviors based on different stages of existing relationships (i.e.presented in the competence formula) between the agents.Concerning future work, we intend to further identify the particular circumstances in which the use of this sophisticated trust model is more relevant.Also, we intend to explore integrity as another dimension of trustworthiness, as well as exploring other ways of combining the trustworthiness dimensions, and to use other antecedents of trust, such as the trustors' own propensity to trust.A lot of future work remains to be done to ensure that the framework is functional in practice.We plan to conduct experiments to evaluate the robustness of the proposed trust mechanism against the badmouthing attacks and non-compliant members.Furthermore, evidential reasoning in the case of conflicting on evidence and evaluate trust based on this method will need to be elaborated and evaluated.Another interesting direction for future work is to develop guidelines in consultation with policy makers to define sounds policy and standard for determining the contract importance based on the risk associated with the proposed methods and the members' preferences.Many different adaptations, tests, and experiments for our model have been left for the future due to lack of time.Future work concerns the a deeper analysis of our proposed model and validates the model with different computational models that proposed by other scholars such as Urbano et al..
[7]ano et al. [7]identified eight dimensions of context {d 1 , d 2 , ..., d 8 } 5 , where dimensions d 1 and d 2 represent the agents trustor and the trustee, respectively; d 3 and d 4 represent the time and the location of agreement; and d 5 , d 6 , d 7 and d 8 identify and characterize the type of the task, its complexity, deadline, and outcome of its realization, respectively.We adapt the Urbano et al.[7]context definition in this paper.For the simplicity, we assume that all agreements performed in the society of agents refer to the same type of task t (d 5 ), although tasks have different degrees of complexity (d 6 ) and deadlines (d 7 ).In this paper, we only consider four dimensions of context, which are d 1 , d 2 , d 5 and d 8 .We exclude time, complexity and deadlines from our formalization.They will be addressed in a future publication. Ubano et al. considered three different outcomes (d 8 ) is defined by O

Table 1 .
Ben (x,y) evaluation for the number of rounds

Table 2 .
Agent x asks different Agent's opinion about agent z in four situations

Table 3 .
Competence evaluation for the agents a, z and y and the number of rounds