A Social Internetworking System (SIS) is the scenario arising when Web users decide to a liate to multiple social networks. Recent studies show an increasing user tendency of creating multiple identities on di erent social systems and exposing, in each of them, di erent traits of their personalities and tastes. This information provides a better picture of user needs and enhances the quality of services they can use. In the next years a large growth of SIS phenomenon is foreseeable. In order to boost the level of user participation in a SIS, suitable mechanisms capable of discerning reliable users must be designed. We propose a model to represent a SIS, a software architecture to gather real data and analyze the structural properties of a SIS. In concrete use cases with di erent contexts and di erent levels of protection of data, we introduce an ontology-based model to compute trust and reputation in a SIS. This research is collaborative e ort between the Vrije Universiteit Amsterdam and the University of Reggio Calabria in the context of a Marie Curie Fellowship.
Social media applications, such as blogs, multimedia and link sharing sites,
question and answering systems, wikis and online forums, are growing at an
unprecedented rate and are estimated to generate a signi cant amount of the contents
currently available on the Web [
In social media applications, users form communities, typically modelled as social networks. Users are driven to get in touch and become friends of other users, create and publish their own contents (like videos or photos), share these contents with others, rate and comment contents posted by others. Examples of popular Web-based social networks are Facebook, MySpace and LinkedIn.
The value of social networks expresses in multiple ways. For instance, users may take an advantage of their interactions with other users to nd information relevant to them or they can explore connections existing in a social network to get in touch with user with whom they may pro tably interact: many Web users, as an example, indicate that they were able to get a job through their contacts in LinkedIn1.
A further advantage is that social networks allow to disseminate new
knowledge in a widespread fashion, to di use innovations, to spread opinions (e.g.,
social or political messages) among members, to advertise new products [
The power of social networks has been fully recognized by institutional actors like museums, TV broadcaster, academic and government institutions. For instance, the Rijksmuseum Amsterdam is exploring the added-value of providing artworks online, allowing users to express their opinions on them or contribute to describing artwork's. Furthermore, major European broadcasters, such as BBC and RAI are experimenting with Web 2.0 technologies to improve interactivity and participation of TV consumers.
Users often decide to a liate to multiple social networks: for instance, in a
recent survey, Ofcom found that 39% of UK adults with at least one social
networking pro le has indeed two or more pro les2. We call Social Internetworking
System (hereafter SIS) the scenario arising when many users decide to a liate
with multiple social networks. Companies are discovering the potential of social
internetworking and are promoting systems capable of supporting social
internetworking tasks. For instance, Google has recently proposed Open Social [
The main goal of these systems is to o er a technological platform to ensure data portability among di erent social networks. The major bottleneck for the success of a SIS is the absence of mechanism that helps users in nding other \reliable" users with whom they can pro tably interact and discloses the presence of malicious users or spammers.
In the past, signi cant research e orts have been done to de ne and handle
trust and reputation, as a large body of literature highlights [
However, in our opinion, several reasons explain a further investigation. A rst research question is to provide a model capable of representing a SIS, its components and their relationships. In addition, it is necessary to gather real data about a SIS in order to understand its structural properties and clarify to what extent a SIS di ers from traditional social networks.
A second issue depends on the fact that the concepts of trust and reputation may assume di erent meanings according to the scenario in which a user operates in. For instance, in communities like Question & Answering systems (in which users are allowed to pose questions, to answer questions raised by other users 1 http://www.mainstreet.com/article/career/employment/social-mediajob-seeker-s-best-new-tool 2 http://www.ofcom.org.uk/advice/media literacy/medlitpub/medlitpubrss/ socialnetworking/annex3.pdf and, nally to rate received answers), the reputation of a user coincides with his level of expertise on a particular topic; in a Web community like YouTube, the reputation a user coincides with the quality of contents he generated. This requires to de ne a procedure to compute, in an abstract and general fashion, the reputation of a user and to specialize it in concrete domains.
As a nal research challenge, it would be necessary to de ne a model to represent trust and reputation in di erent contexts. In addition, it is useful to observe that, in di erent contexts, di erent policies for accessing, publishing and re-distributing data may exist. For instance, in the case of a TV broadcaster which delivers online part of its archive of resources, users are allowed for instance to use some resources (e.g., for educational purposes) but are forbidden to re-use those protected by copyright. To address these issues, it would be bene cial to design an ontology to model the key concepts of trust and reputation in di erent environments characterized by di erent levels of protection of data.
In this paper we propose a methodology to handle trust and reputation in a
SIS. The paper summarizes the research activities we are planning to carry out
in the context of a Marie Curie Intra-European Fellowships for Career
Development (IEF), a funding opportunity provided by EU Commission. The paper
is structured as follows: in Section 2 we review existing trust and reputation
models and illustrate the challenges arising in a SIS. In Section 3 we illustrate
a model to represent the features of a SIS along with a software architecture
we are currently implementing to gather real data from a SIS and analyze its
structural properties. In Section 4 we provide a general model to compute trust
and reputation in a SIS and illustrate the steps we are planning to specialize it
in real contexts; in particular, we plan to adapt our notion of trust on data
gathered within two research projects, namely NoTube [
In virtual communities the term trust is generally exploited to indicate the reliance that a community member associates with another one. Trust values are \local parameters" in the sense that specifying the trust of a user A toward a user B is equivalent to indicate how much A perceives B as reliable.
The opinion of the whole community of users toward a member of the community itself is known as reputation. In the past, the issue of computing and handling trust and reputation in virtual communities has been deeply investigated and several models and approaches to facing it have been proposed.
Here we discuss some of these approaches and outline the challenges we encounter in the context of a SIS. Existing approaches can be classi ed into two categories:
Graph-Based Approaches. A rst category of approaches model a user
community as a graph G in which nodes represent users [
Graph-based approaches leverage on explicit trust relationships declared between pairs of users. As a consequence, they neglect to consider a broad range of activities that, in a SIS (e.g., the activity of rating resources) are a precious and reliable indicator of trust.
Link-Based approaches. A second category of approaches use ranking
algorithms such as PageRank [
Experimental tests indicated that link-based methods can obtain precise results and are often attack-resistant, i.e., they can resist to attempts conceived to manipulate reputation scores.
We observe that in some approaches trust is conceived as a measure of
performance. For instance, in [
We can observe that both graph-based and link-based approaches try to model trust and reputation in a \force-mass-acceleration" style. In other words, these approaches try to capture all factors in uencing trust and reputation and combine them in a set of equations. The resulting model is too complicated to be handled and it may not provide signi cant results. In our opinion, the assessment of trust and reputation critically depends on the concrete domain in which we are operating in and we believe that an universal model of trust is not possible.
To better clarify this concept, we report some results emerging from the
PrestoPrime project [
In the context of PrestoPrime, two pilot demonstrators were developed. In the rst one, in a game-like environment, users were asked to label videos by applying simple keywords (tags). Experiments with users showed that a satisfactory measure of trust between a pair of users who do not know each other can be obtained by considering the tags they apply to label a video and computing the degree of match of the set of tags they inserted.
In the second demonstrator, users were provided with a small annotation environment allowing them to label museum objects with four main entry elds (i.e., \who", \where", \what", and \when"). This allowed us to create links between museum objects on the basis of the key dimensions \who", \where", \what", and \when"; as an example, objects coming from di erent museum collections can be tied if they refer to the same artistic and historical context and this produces a more complete description of cultural movements.
The notion of trust developed in the context of the rst demonstrator is not applicable for the second one, and other factors in uencing trust and reputation need to be studied.
A further challenge we are in charge of studying depends on the fact that, in some cases, real organizations often decide to make available on the Web their own resources and often allow end users to enrich their descriptions through metadata like tags. For instance, think of the case of public TV broadcasters like BBC which o ers online a large number of contents referring to its TV programmes. Each organization may use di erent policies for accessing, distributing and labelling the contents they produce and disseminate. For instance, a digital content may published online only in some speci c cases (e.g., if the material must be used in education) while its usage is forbidden for commercial purposes. This proves that, in the process of de ning trust and reputation, it is necessary to consider not only the application context but also te level of data protection about available resources. 3
The rst goal of our research is to nd a suitable model to represent a Social Internetworking System (SIS) and interactions between humans that can take place in it.
To this purpose, our model must t two requirements: { Requirement 1. The model should be rich enough to represent a wide range of heterogeneous entities (i.e,. users, resources, posts, comments, ratings, and so on) and their interactions (e.g., users may declare to be friends or they may rate resources). { Requirement 2. The model should be easy to manipulate and intuitive.
Clearly, Requirements 1 and 2 are con icting each other and a suitable tradeo is compulsory. Traditional approaches to modelling social networks are usually based on graphs. Nodes in graphs represent social network actors (e.g., users) while edges identify relationships between them.
We believe that graph-based models are not satisfactory in the context of a SIS for several reasons.
A rst weakness relies on the role the nodes would have if we would decide to represent a SIS through a graph. Generally, a social network consists of homogeneous nodes, i.e., all nodes represent objects sharing the same nature. As claimed by Requirement 1, in a SIS heterogeneous entities co-exist and these heterogeneities must be properly modelled.
A further limitation is that graph-based models are able to represent onedimensional networks, i.e., edges of a graph specify that only one particular kind of relationship may exist between nodes. On the contrary, we expect that a SIS should be represented through a multi-dimensional network because various type of interactions may involve entities of the same type or of di erent nature: for instance, an edge should link a user u and a resource r to indicate that u has posted r or an edge should tie two users to indicate that they declared to be friends.
Finally, edges in graphs highlight binary relationships between nodes they link. In a SIS, it could be useful to consider n-ary relationships (e.g., an edge may glue together a user u, a resource r and a tag t under the hypothesis that u applied t to label r).
We are currently studying a more sophisticated model in which a SIS is represented through an hypergraph such that: (i) nodes are labelled and the label of a node re ects the nature of the object represented by the node itself; (ii) multiple hyperedges may run between two nodes to indicate that multiple interactions may take place between two arbitrary entities; (iii) hyperedges denote relationships involving two or more entities.
In addition to de ning a model to represent a SIS, we are also interested in gathering data from real social networks in order to understand the properties showed by a real SIS. For instance, it would be interesting to check whether properties typical of real social networks (e.g., the small world phenomenon) still emerge in a SIS.
Such a task is quite complex because, in di erent networks, a user may have di erent identities so it would be extremely hard to join information scattered across multiple networks.
To address this issue, we used the Google Social Graph API [
As a further step, we are interested in studying a model of reputation for a SIS. The notion of trust/reputation in the context of a SIS (and, in general, for Web 2.0 applications) is hard to de ne because, as shown in Section 2, in di erent contexts, they may assume di erent meanings.
We propose a methodology to compute reputation in a SIS which operate in two stages: in the former stage we review and analyze the factors capable of in uencing the value of reputation in a SIS. In the latter stage we consider some concrete domains and specialize our methodology to them.
Intuitively, we assume that the reputation of a user depends on the following facts:
The reputation of a user depends on the relationship he created in the SIS. We suggest to use the hypergraph model introduced in Section 3 to represent users and their relationships in a SIS. Past user interactions are analyzed to determine the level of trust a user u confers to a user v and this information is used to weight edges in the hypergraph representing the SIS. Since a user generally interact with an handful of other users, the hypergraph we obtain is sparse and a suitable algorithm to propagate trust values is necessary. Currently, we are planning to use a link-based algorithm like PageRank. At the end of this step, we are able to generate a vector r0 such that the i-th component of r0 equals to the reputation of the i-th user.
Users with high level of reputation are also those who produce high quality resources. The quality of a resource could be computed by consider the average rating it got and, then, resources with a high average rating are also high quality resources. To avoid biases, we can pose a further requirement: the number of ratings received by a resource must be statistically signi cant, i.e., we can consider only resources which received at least Nmin ratings, being Nmin a suitable threshold. Such a requirement would avoid that resources evaluated by a small number of users are deemed better than resources rated by a large mass of human users.
The procedure described above resembles that applied in many social systems like YouTube or Digg to evaluate the quality of a resource. We believe that such a procedure is a ected by several fallacies and it may incur in harsh inaccuracies. In fact, spam or malicious users may tend to provide generous evaluations to arti cially in ate the evaluation of a resource. As a consequence, we need a more complicated framework capable of putting together the reputation of users, the quality of resources they post and the evaluations associated with resources. At the current stage of the project we are considering, as a possible solution, the following criterium:
A user has a high reputation if he authors high quality resources. A resource, in its turn, is of high quality if it gets a high average rating and it has been posted by users with high reputation.
The intuition provided above relies on a mutual reinforcement principle that
is similar, to some extent, the approach underlying HITS [
According to this notation, we can write the following equations: (1) (2) 1 r00 _ Aq q _ AT r00 + e
In both Equations 1 and 2, the symbol _ means \is proportional to". As for Equation 1, the i-th row of the the product Aq speci es the sum of the qualities of the resources authored by the i-th user. This immediately follows from the de nition of product between a matrix and a vector. Interestingly enough, the A matrix can be interpreted as the adjacency matrix of a bipartite graph whose nodes represent users and resources and edges link a user to the resources he authored. In Equation 2, the symbol AT is the transpose of A. As in the previous case, AT matrix can be viewed as the adjacency matrix of a bipartite graph whose nodes represent resources and users and edges link a resource to the user who authored it. Observe that the same model holds if we assume that a resource has been posted by one user or it has been posted by multiple users. The product AT r00 is an m-th dimensional vector whose j-th entry speci es the reputation (or the sum of the reputations) of the user (users) who posted the j-th resource.
By plugging Equation 2 into Equation 1 we obtain: r00 _ A hAT r00 + ei ) r00
AAT r00 _ Ae ) ) r 00
I
AAT _ Ae ) r00 _ I
AAT 1
Ae
Since AAT is symmetric, its eigenvalues are real [
Finally, we merge the arrays r0 and r into a single reputation value r as follows: r = r0 + (1 ) r 00 (3)
The coe cient 2 [0; 1] is instrumental in weighting the contributions coming from link analysis and the analysis of resources generated by a user. We plan to tune by applying a linear regression technique.
Once a theoretical model of reputation in a SIS has been de ned, our intention is to specialize it in concrete domains. In particular, we are interested in monitoring and analyzing the behaviour of users in long-term experiments associated with di erent domains; the notion of reputation, from abstract concept turns into a concrete tool to aid user in better taking advantage of potentialities o ered by the SIS. Experiments on real users allow us to get an iterative assessment of the strengths and weaknesses of our notion of reputation as well as indications for improvement.
To this purpose, we will use data gathered in the context of two research
projects, namely: (i) NoTube (an EU nanced project on interactive television)
[
Once we have de ned the concept of trust and reputation in concrete domains it is advantageous to create a model capable of representing them in di erent domains. To this purpose we plan to design an ontology capable of specifying how reputation and trust specialize in di erent application contexts.
To the better of our knowledge, there are few approaches to designing
ontologies to model trust. For instance, in [
Our goal is di erent from that of [
The availability of such an ontology would o er us, on the long run, the possibility of designing complex software applications running across multiple social networks. For instance, we can think of a content-based recommender system operating as follows: 1. A user issues a query. 2. The query is forwarded to multiple social system and a list of resources matching the query is retrieved by each social system. 3. Retrieved resources are ranked on the basis of the reputation of the users who created, on the application context and on the rights for its distribution. 4. A global list is produced by merging the previous ones.
Such an application is, in our opinion, capable of introducing relevant novelties in the research eld of Recommender Systems. In fact, the proposed application is able to sift through di erent social sites (while traditional Recommender Systems usually operate on a single resource repository) and is able to rank resources on the basis of multiple and criteria. 6
In this paper we introduce the concept of Social Internetworking System, i.e., the scenario arising when Web users decide to a liate to multiple social networks. We propose a model to represent a SIS and describe the main components of a software architecture we are implementing to gather real data from a SIS and analyze its structural properties. In concrete use cases with di erent contexts and di erent levels of protection of data, we introduced an ontology-based model to compute trust and reputation in a SIS. This research is collaborative e ort between the Vrije Universiteit Amsterdam and the University of Reggio Calabria in the context of a Marie Curie Fellowship.
In the future we plan to gather a large amount of data about a SIS and carry out an empirical study on them. The goal is to understand whether some properties of real social networks (like small world phenomenon, power law distribution of in-degree and out-degree distributions, and so on) if they are still con rmed in a SIS or if signi cant deviations emerge.
A further research line is to carry out a detailed review of existing literature on the meaning of trust and reputation in di erent social site. Finally, we plan to test the e ectiveness of our ontology-based model with an experiment involving real users. In particular, the validation phase will be strictly tied to the activity of designing our ontology; in fact, we shall use feedbacks provided by users to revise the structure of our ontology.