Final Report Summary - TRUSIS (Trust in Social Internetworking Systems)
On the Social Web that now is omnipresent in all of our daily lives, many people have accounts in different social spaces. They might have a profile on Facebook to share with their friends their daily activities, they might have an account on YouTube to upload, share and view video clips, they might have an account on Flickr to post, share and view photos, and they might have many more of those profiles and accounts in different social network applications. The TRUSIS research project has concentrated on the problem of understanding the behaviour of users that are affiliated with multiple social networks. Whereas other research has typically concentrated on single social networks, TRUSIS has started with the observation that today many users create an account in different social spaces (e.g. Facebook, YouTube, Flickr) and that the digital identities of the users are not interlinked with each other but appear as isolated entities.
This observation implies that the different participations of people in social networks are perceived as independent and separate, and as a consequence our potential to obtain insight in what users are doing in social spaces is limited to the scope of single, separate social networks. If however we would able to effectively glue those social networks together for analytical purposes, we could obtain a more complete understanding of user aptitudes and needs. Such an improved understanding of users is an effective basis for software developers that allows to design software applications capable of better satisfying user needs. In summary, the main challenge of TRUSIS was to allow to design better personalized software by analyzing the interlinking of social network systems.
The TRUSIS project has approached its main challenge in several steps.
As a first and initial step in the research activities, we have concentrated on the analysis of user behaviours in multiple social spaces. This behaviour of users in multiple social spaces was studied in terms of the relationships they create. The goal was to identify how the relationship creation plays out in multiple social networks. For this purpose, we built a crawler that was based on the Google Social Graph API, a tool recently released by Google and capable of identifying connections between users on the Web. The crawler allowed us to retrieve information about 1.3 millions of user accounts throughout several social networks (like Flickr, LiveJournal and Twitter); it also allowed us to retrieve 36.2 millions of connections among these accounts. We subsequently mapped the data thus gathered onto a graph and used techniques for Social Network Analysis to analyze the interrelated social networks.
From this analysis, we could derive the following observations to hold about user behavior in multiple social networks:
(i) the distribution of user contacts follows a power law;
(ii) there exist close-knit clusters of users that are densely connected to each other;
(iii) there exist some mediators, i.e. users who are important in two or more social networks and foster information circulation across them.
To exploit these findings and create socio-economic benefits we considered recommendation algorithms. The topological analysis of user connections we did in the first steps, allowed us to then design a new algorithm to recommend users with new social contacts. In detail, we suggested to use a popular coefficient from the field of sociometry that is known as the Katz coefficient to determine the similarity degree of two users on the basis of their social contacts. We provided also an efficient algorithmic framework to quickly approximate the values of the Katz coefficient, an algorithm that performed well even in the presence of very large networks.
As a further step, building on the promising outcomes of the first steps, we turned our attention to a further analysis of user behaviors, in which not only topological information was considered but also information at the semantic level. We were aiming to investigate again the interlinking between multiple social networks but this time also considering semantic knowledge about the social network content and metadata. In detail, we used a dataset consisting of about 432 000 user profiles gathered from the Web through the Google Social Graph API. In this dataset, we isolated users who simultaneously had an account in Delicious, StumbleUpon and Flickr.
Subsequently, for each selected user we extracted the list of her social contacts as well as the tags she exploited. We then were able to study whether user behaviour was homogeneous across different social systems: we could see if users who tend to use few tags in one system tend to use also few tags in another one. We could also study whether a form of correlation exists between tagging and social activities like friending: we could investigate whether users who are frequently involved in tagging activities are also socially active. Finally, we created opportunities to use Semantic Web technologies (including the use of the DBpedia interface) to map user tags onto categories and to determine to what extent tagging activities in different social systems are biased toward a specific topic and determine whether a user focuses on the same topics despite that she is active in different social systems.
The application benefits of this analytical knowledge are in the specific targeting of personalization towards the social network system where the user is best served for a particular topic: the socio-economic benefits of better personalization are obvious, both for the people using the social networks as well as the private and public bodies offering services that can benefit from this enhanced personalization.
As a third research line, we concentrated on the computation of trust in cross social systems. In this research line, we focused on the problem of extending the concept of user trust in a social network to an ensemble of social systems. While the understanding of trust in a single social network is beginning to mature, the definition of trust in connected social systems is still in a phase of infancy and therefore the potential impact of our attempts to define it are immense. A first research action was devoted to study what requirements are useful to define trust and how different factors (like social connections, tagging and rating activity, etc.) impact the process of trust computation.
As a second step in this research line, we studied how to concretely define a trust model in two or more social networks. We investigated how to use such a model to generate recommendations. In detail, we first introduced the concept of social evaluation networks, i.e. social networks in which users are allowed to perform various evaluation activities on the available contents, like rating contents or posting comments or reviews. These evaluation activities then were translated into a matrix that was storing the numerical evaluations the users provided on the items. We then proposed a matrix-based factorization algorithm to alleviate data sparsity in the evaluation matrix.
The proposed algorithm differs from the current algorithms in the sense that
(i) information about social relationships between users are incorporated in the matrix factorization process and
(ii) data available in a given social network can be used in another one to raise the quality of information.
First experiments have been run and more experiments are being are now being run on both a dataset crawled on the Web, from epinions.com and data that was manually collected from a community of students at University Mediterranea. The results point towards an increased understanding of trust cross social networks, which translates immediately in benefits for private and public service providers and clear societal benefits of a more coherent and trusted society.
These research activities were carried out by researchers from VU University in Amsterdam, the Netherlands, and from University Mediterranea of Reggio Calabria, Italy, in conjunction with researchers from Delft University of Technology in Delft, the Netherlands, and University of Turin, Italy.
This observation implies that the different participations of people in social networks are perceived as independent and separate, and as a consequence our potential to obtain insight in what users are doing in social spaces is limited to the scope of single, separate social networks. If however we would able to effectively glue those social networks together for analytical purposes, we could obtain a more complete understanding of user aptitudes and needs. Such an improved understanding of users is an effective basis for software developers that allows to design software applications capable of better satisfying user needs. In summary, the main challenge of TRUSIS was to allow to design better personalized software by analyzing the interlinking of social network systems.
The TRUSIS project has approached its main challenge in several steps.
As a first and initial step in the research activities, we have concentrated on the analysis of user behaviours in multiple social spaces. This behaviour of users in multiple social spaces was studied in terms of the relationships they create. The goal was to identify how the relationship creation plays out in multiple social networks. For this purpose, we built a crawler that was based on the Google Social Graph API, a tool recently released by Google and capable of identifying connections between users on the Web. The crawler allowed us to retrieve information about 1.3 millions of user accounts throughout several social networks (like Flickr, LiveJournal and Twitter); it also allowed us to retrieve 36.2 millions of connections among these accounts. We subsequently mapped the data thus gathered onto a graph and used techniques for Social Network Analysis to analyze the interrelated social networks.
From this analysis, we could derive the following observations to hold about user behavior in multiple social networks:
(i) the distribution of user contacts follows a power law;
(ii) there exist close-knit clusters of users that are densely connected to each other;
(iii) there exist some mediators, i.e. users who are important in two or more social networks and foster information circulation across them.
To exploit these findings and create socio-economic benefits we considered recommendation algorithms. The topological analysis of user connections we did in the first steps, allowed us to then design a new algorithm to recommend users with new social contacts. In detail, we suggested to use a popular coefficient from the field of sociometry that is known as the Katz coefficient to determine the similarity degree of two users on the basis of their social contacts. We provided also an efficient algorithmic framework to quickly approximate the values of the Katz coefficient, an algorithm that performed well even in the presence of very large networks.
As a further step, building on the promising outcomes of the first steps, we turned our attention to a further analysis of user behaviors, in which not only topological information was considered but also information at the semantic level. We were aiming to investigate again the interlinking between multiple social networks but this time also considering semantic knowledge about the social network content and metadata. In detail, we used a dataset consisting of about 432 000 user profiles gathered from the Web through the Google Social Graph API. In this dataset, we isolated users who simultaneously had an account in Delicious, StumbleUpon and Flickr.
Subsequently, for each selected user we extracted the list of her social contacts as well as the tags she exploited. We then were able to study whether user behaviour was homogeneous across different social systems: we could see if users who tend to use few tags in one system tend to use also few tags in another one. We could also study whether a form of correlation exists between tagging and social activities like friending: we could investigate whether users who are frequently involved in tagging activities are also socially active. Finally, we created opportunities to use Semantic Web technologies (including the use of the DBpedia interface) to map user tags onto categories and to determine to what extent tagging activities in different social systems are biased toward a specific topic and determine whether a user focuses on the same topics despite that she is active in different social systems.
The application benefits of this analytical knowledge are in the specific targeting of personalization towards the social network system where the user is best served for a particular topic: the socio-economic benefits of better personalization are obvious, both for the people using the social networks as well as the private and public bodies offering services that can benefit from this enhanced personalization.
As a third research line, we concentrated on the computation of trust in cross social systems. In this research line, we focused on the problem of extending the concept of user trust in a social network to an ensemble of social systems. While the understanding of trust in a single social network is beginning to mature, the definition of trust in connected social systems is still in a phase of infancy and therefore the potential impact of our attempts to define it are immense. A first research action was devoted to study what requirements are useful to define trust and how different factors (like social connections, tagging and rating activity, etc.) impact the process of trust computation.
As a second step in this research line, we studied how to concretely define a trust model in two or more social networks. We investigated how to use such a model to generate recommendations. In detail, we first introduced the concept of social evaluation networks, i.e. social networks in which users are allowed to perform various evaluation activities on the available contents, like rating contents or posting comments or reviews. These evaluation activities then were translated into a matrix that was storing the numerical evaluations the users provided on the items. We then proposed a matrix-based factorization algorithm to alleviate data sparsity in the evaluation matrix.
The proposed algorithm differs from the current algorithms in the sense that
(i) information about social relationships between users are incorporated in the matrix factorization process and
(ii) data available in a given social network can be used in another one to raise the quality of information.
First experiments have been run and more experiments are being are now being run on both a dataset crawled on the Web, from epinions.com and data that was manually collected from a community of students at University Mediterranea. The results point towards an increased understanding of trust cross social networks, which translates immediately in benefits for private and public service providers and clear societal benefits of a more coherent and trusted society.
These research activities were carried out by researchers from VU University in Amsterdam, the Netherlands, and from University Mediterranea of Reggio Calabria, Italy, in conjunction with researchers from Delft University of Technology in Delft, the Netherlands, and University of Turin, Italy.