-Evidence suggests that people often waver to buy trust is represented by a number of models that rely on global from online vendors because of uncertainty about vendor behav- reputation [4]-[6]: they are based on the evaluation of the ior or the risk of having wrong information about the products. behaviors of the users, that is shared across the entire commupTerrucsetptpi olanyss oaf rcisekn.trMalorreooleveri,nthheemlpaitnicg gcroonuspusmaerres goavienricnogmae nity. These models, however, show an evident limitation due lot of attention and high centrality in online community, as users to the difficulty of taking the effects of malicious or fraudulent share opinions and/or mutually collaborate for reaching their behaviors into account, thus making the feedback themselves. targets. The users can be helped by personal software agents Other approaches, that consider also a local perspective of able to perform activities aimed at supporting the purchase of the trust, are limited by the fact they are supervised, i.e. nperotwduorcktss. Torhgisanpizaepderbpyrogproosuepss.a Innewpatrrtuicsutlamr,eawsuereprinesesnotciaal they need a training phase in generating recommendations. model to represent this scenario, and we introduce an algorithm In [7], we proposed to integrate the traditional use of the for detecting trust recommendations in virtual communities in global reputation with the local reputation, that is based on presence of groups. We technically formalize our idea and show the recommendations coming by the entourage of the user a complete example of how our approach works. (friends, friends of friends and so on). But this proposal was Index Terms-Recommendation, Online Communities, Trust, limited because it does not consider a group-based structure Group. [8]-[10]. In this paper, we define a new model to represent the I. INTRODUCTION groups of users linked by trust relationships. Such a model depends on three main parameters: the relevance given to the reliability with respect to the reputation, the threshold of recommendation under which a product can be considered as not interesting and the number of the groups in online communities. We propose an algorithm for detecting trust recommendations for a user considering the recommendations that come from users within his own group and those of other groups weighted with the global reputation. We technically formalize our idea and algorithm, and we present a complete example of how our approach works. The paper is organized as follows: in Section II we deal with some related work. Section III provides technical details about our approach for finding trust recommendations about products, while Section IV describes a concrete example of application. Finally, in Section V we draw our conclusions and illustrate some possible future works.
An important issue in Online Social Networks (OSNs) is
that of designing recommender systems capable to provide
OSN users with useful suggestions regarding other potentially
promising OSN users to contact as interlocutors or interesting
content to access. Such an issue leads to the necessity of
considering the opinions that different users express about other
users or OSN content [
The issue of trusting own interlocutors widely emerged in large online e-Commerce communities as, for instance eBay, and now it is largely discussed in many OSNs which allow their users to create and share contents with other users as well as opinions. This is the case, for example, of well-focused OSNs like EPINIONS1 and CIAO2, in which users provide reviews concerning commercial products falling in different categories. Almost all of these platforms face this issue by adopting a reputation system. Reputation is a form of indirect trust, where a user takes advantage from the opinions coming from other users for evaluating the probable trustworthiness of an interlocutor. Commonly, in the traditional OSN contexts, the reputation of a user is evaluated by averaging feedbacks provided by all the other users belonging to the same community. In the past literature, a common approach for predicting
A large number of papers in the literature investigated on the topic we deal with here, therefore, in this section we cite only those approaches which we consider comparable with that discussed in this paper.
Concerning the concept of trust, there exist in the literature
several proposals. Sherchan et al. [
In many disciplines, there is a population of people which
should be optimally divided into multiple groups based on
certain attributes to collaboratively perform a particular task
[
Our scenario is represented by a virtual community S,
formally denoted as S = ⟨A, G⟩, where A is the set of agents
joined with S and G is the set of groups contained in S. We
also assume that each group g is managed by an administrator
agent ag. Generally, all such communities are organized in
social structures based on social relationship (like, Facebook
[
The trust measure tu,v is a mapping that receives as input two agents u and v and yields as output a boolean value representing the degree of trust between two agents u and v: tu,v = 0 (resp. tu,v = 1) means that u assigns the minimum (resp. maximum) trustworthiness to v. The trust measure is asymmetric, in the sense that we do not automatically expect that v trusts u at the same level.
As a theoretical proposal, we had introduced a more general trust measure, by combining two components relu,v and repu, where (i) relu,v is the direct reliability of u, i.e. the trustworthiness that v has in u based on the past interactions between u and v while (ii) repu is the global reputation of (1) (2) (3) u, i.e. the trustworthiness that all the community has in u. The reason of this choice, was due to the necessity, when v does not have a sufficient direct knowledge of u, to use the recommendations coming from the other agents of the community.
1) Reliability: As for the reliability, we denote it by the mapping relu,v, assuming values ranging in the domain [0 · · · 1] ∪ N U LL, while relu,v = NULL means that v did not have past interactions with u and thus it is not able to evaluate us’ trustworthiness.
2) Reputation: As for the reputation of u, we denote it by repu in the interval [0 · · · 1]ϵR. In order to compute the reputation, we adopt the notion of repu = 1
∑ hmax| REVu| ρ∈REVu hρ where | REVu| is the set of the reviews made by the user u and h is the helpfulness, i.e., it is associated with each review that represents the level of satisfaction of the other users for that review. To normalize repu, we divide it by the maximum value of the helpfulness hmax.
The two trust components reliability and reputation are integrated in a unique value to compute the mapping trust tu,v of u about v, producing a input ranging in [0 · · · 1] as follows:
tu,v = α · relu,v + (1 − α) · repv where α is a real number, ranging in [0...1], which is set by u to weight the relevance he/she assigns to the reliability with respect to the reputation.
The user receives, at the current step, some recommendations about the products present in the community. In other words, recpu is the recommendation that the user u receives about the product p. It is calculated as follows: recpu = ∑v∈ρ,v̸=u tu,v · ratevp
∑v∈ρ,v̸=u tu,v
where ratevp is the review of the the user v about the product p (a number between 1 and 6), weighed by the trust of v. This means that his/her opinion about a product is taken into account if his/her trustworthiness is high. The weighted average allows us to identify an average value in which the starting numerical values have their own importance, specified by its weight. In particular, we can identify the center of gravity of the rate. In this way, we give more importance to the rate from users that the user u trusts. With a normal mean we would lose significant information.
At this point, we introduce the groups’ concept in the community. In this context, we define trust t∗u,v in two different ways. We suppose that the trust perceived by an agent u with respect to the component of his/her group is equal to 1 (i.e, t∗u,v=1), instead the agent u considers the trust that user has in the whole community (see Equation 2). In this way, we define recg∗p that is the recommendation that the user u receives about the product p in presence of the groups: rec∗up = ∑v∈gu ratevp + ∑v∈/gu tu,v · ratevp ∑v∈| REVp| t∗u,v (4) where gu is the group to which the agent u belongs and | REVp| is the set of the agents who have purchased the product p. It is calculated as the combination of two contribution: the average rating of the users that belong to the group of the user u and the score that the other groups give to the product multiplied by the trust that u assign to its agents.
Now, we explain how it is possible to use groups to generate the recommendations of products for the users inserted into an online e-Commerce communities. As an example, we propose to model each user by a node (see Figure 1).
We assume that all the elements of a group are trust-related, a trust group g determined into G represents a mutual trust relationship between its elements. In our case, there are four groups of users called g1, g2, g3 and g4. All the users in the same group are mutually linked by a trust relationship with the value 1; while the values of the reliability are shown in the Table I.
In particular, in our example, there are nine products that are divided in three main categories called Electronics, Informatics and Software (see Table II). In the Table III, we show an example of datasets.
In our model, we have associated with each agent a profile contained, as unique feature, the reputation to review the products. This reputation has been computed by averaging, on all the reviews posted by the agent, the helpfulness associated with each review, where the helpfulness is an information available on the dataset and obtained by the opinions expressed by the users of the community. We obtain that the reputation values (see Equation 1) of the agents belonging to our scenario are as follows: rep1=0.83; rep2=0.75; rep3=0.66; rep4=0.41; rep5=1; rep6=0.62; rep7=0.45; rep8=0.5; rep9=0.79; rep10=0 and rep11=0.58. At this point, we introduce a new value of trust tu,v that is the combination of the reliability and the reputation. Fixed the agent a1, we compute the opinion (i.e., trust) that a1 has with regard to other agents. Recall that this value changes with α. We consider three values of α. In particular, α=1 means that the agent a1 considers only the opinions of the agents with whom he/she interacted in the past (contrariwise, α=0). Finally, α=0.5 means that the agent a1 considers in the same way both the opinions of the agents with whom he/she interacted both others. For detail, see Tables V- VII. Let ξ be a threshold fixed by the agent a1, we suggest only those products that have recpu greater than ξ (in our case, we fix ξ > 4). In particular, we note that the agent a5 that has a high value of reliability for the agent a1 buys the products p8 and p9. Also, u5 assigns to them a high rate while the rest of the community gives a very low rate. Surely a1 would be very interested in these products, because he/she trusts a5 with a high value. At this point, we see how the algorithm behaves. The agent a1 receives, at the current step, some recommendations by the other agents, in response to previous recommendation requests (see Table IV). If α=1, we suggest to a1 the products p6, p8 and p9. In this case, we consider the recommendations that come from agents that have a high reliability. In fact, these products were acquired and evaluated good by agents a5 and a2. Comparing these results with truly user-purchased products, it is visible that three out of three products were actually purchased by u1. If α=0, we suggest to a1 the products p1, p4 and p5. This is correct because they are products purchased by agents who have a high reputation in the community. But, these agents did not interact directly with a1 therefore they do not know his/her preferences. Indeed, only the product p1 is of interest to a1. If α=0.5, we suggest to a1 the products p4, p5, p8 and p9. This choice is a good compromise, since three of the four products are of liking for the agent a1.
With the introduction of the groups in the community, we can consider the assumption made in the Section III. Recall that, for the agents that are in the same group, the trust is equal to 1. In our case, a1 is in the group g1 with the agents
TABLE VI SIMULATION PARAMETERS FOR α=0.5 a3 and a7, therefore t1∗3 and t1∗7 are always equal to 1. Now, we can calculate the recommendations (see Equation 4) to the agent a1 for all the products in the community in presence of the groups. The table VIII shows the results obtained.
The results in the presence of the groups are best, because a3 and a7 know better a1 and consequently are able to make targeted recommendations.
In order to model the process of evaluating the performance of recommendation provided by learning agents, we defined two indexes. Let be Ri the set of recommendations provided to the agents ai, and Ri ⊂ Ri the set of recommendations relating to the products which ai purchased. Besides, let be Γi∗ the set of purchases made by ai, where Γi is the set of all actions executed within the context of agent ai.
To provide a performance measure we defined two indexes, called Precision and Recall, as follows:
P re(Ri) = | Γi∗ ∩ Ri|
| Ri| Rec(Ri) = | Γi∗ ∩ Ri| | Γi∗| (5) (6)
By the definition of P re(Ri), it follows that a high value does not mean to be a good recommender agent. Indeed, it is possible that the overall performances of provided recommendations are not the greatest possible. Rec(Ri), which is the fraction of recommendations successfully suggested by the agent ai, allows us to consider the aspect above. In our case, we have different values of Precision and Recall to vary by α (see Table IX). It is clear that when ai takes into account only the opinion of the whole community, we have relatively low performances. Instead, combining the opinion of the whole community with that of the agents who had direct interactions with ai, we obtain high values of Precision and Recall. In particular, when α=0.5 in the absence of the groups we have a higher value of Recall because the agents who belong to the community but had no interactions with ai bought many products and then their evaluations are appreciated in the community. This situation allows to advise the products that are of interest for ai. However, when α=1 in the presence of the groups, we obtain Rec(Ri)=0.8 that is the highest value. This means that the recommendations of agents within the group joined to those of the agents with which ai has interacted in the past, allow us to suggest the products of his/her interest with very high accuracy (88%).
In this paper, we propose a model capable to integrate reliability and reputation in an OSN organized by groups. In particular, we considered three important parameters in order to characterize the model: the relevance given to the reliability with respect to the reputation, the threshold of recommendation under which a product can be considered as not interesting and the number of the groups. We have presented a realistic example and the results have shown that when the agent takes into account only the reputation, we have low performances. Instead, combining the opinion of the whole community (reputation) with that of the agents who had direct interactions with her/him (reliability), we obtain high values of Precision and Recall. However, in the presence of the groups, we obtain that Recall has the highest value. In other words, in this letter case, our model allows to suggest the products with very high accuracy (88%). In this paper, we limited ourselves to introduce and formalize the idea, and we present an example of how the presented approach can found product recommendations in an online e-Commerce communities. Our ongoing research is currently devoted to apply the approach to real social networks, in which the advantages and limitations introduced by our proposal can be quantitatively and effectively evaluated.