Nowadays, users join several on-line social networks (OSNs) so that design and development applications able to work on multiple OSNs is a challenging issue. However, OSNs present relevant di erences from both the adopted terminology (similar concepts have di erent names) and the supported technology (for example, in the APIs provided for data extraction). Consequently, the heterogeneity of OSNs does not allow the design of applications with suitable abstraction with respect to the speci c OSNs processed. In this paper, we de ne a model aimed at generalizing concepts, actions and relationships of existing social networks, which can be exploited as a middleware to implement applications working on multiple social networks.
Over the past decade, online social networks have became part of people's live.
Nowadays, most people have a pro le in one or more online social networks like
Facebook, Twitter, Linkedin, MySpace, in which they spend a lot of time. This
is recognized as an important phenomenon from a social and economic point
of view, and, thus, in design and development processes of (Web) applications.
Indeed, often applications should be based on behaviors of a community, or take
advantage from these, so that modern Web applications should be social by
default. In many cases, both personal information and social interactions coming
from social network pro les can be part of innovative solutions. Among these,
social Web applications are the most signi cant example, in which both people's
identities and contents they produced are involved in the business process and
data are mostly owned by users, strongly interlinked and inherently polymorphic
? This is a short version of the paper titled \A model to support design and
development of multiple-social-network applications" [3], which appears in Information
Science journal.
[
On the other hand, the power of the social-network substrate can be fully exploited only if we move from a single-social-network to a multiple-social-network perspective, still keeping the user-centered vision, so that the above issue becomes crucial. The recent literature has highlighted that the aforementioned multiple-social-network perspective opens a lot of new problems in terms of analysis [12] but also new opportunities from the application point of view [14, 8, 15, 18]. Even though each single social network is an extraordinary source of knowledge, the information power of the social-network Web can be considerable increased if we see it as a huge global social network, composed of autonomous components with strong correlation and interaction. Thus, social-network-based programming should work at this abstraction level.
In this paper, we do an important step to cover the gap highlighted above, by de ning and implementing a model aimed at generalizing concepts, actions and relationships of existing social networks. This paper is organized as follows. Section 2 introduces the characteristics of the multiple-social-network scenario that we model. We give a formal de nition of the graph-based conceptual model in Section 3. To validate our approach, in Section 4, we show how our model has been pro tably applied to two very relevant applications in the context of social network analysis. Finally, our conclusions are summarized in Section 5. 2
One of the motivation of this study is the strong heterogeneity in the representation of concepts among di erent social networks. For instance, contacts are represented by friends in Facebook and the relationship is symmetric, while they are represented by followers and followings in Twitter and the corresponding relationship is not symmetric. Again, the concept of appreciation becomes +1 in Google+ and endorsement in about.me. Importantly, similar concepts can mapped to each other but they have in general di erent features. Thus, an integration step is necessary for our purpose. In this section, we prepare this integration step by grouping the main technical entities into a number of categories to which the formal model presented in the next section maps. In particular, we aim at modeling the following entities.
Pro le. Social network sites are built around user pro les, a form of individual (or group) homepage, which provides a description of each registered user. For example, in Twitter, at the moment of registration, a user can create his pro le typing his name, username, password and email address in the registration form. After, he can upload a pro le picture and start following other people. Moreover he can complete his pro le adding a short biography, a position (the place where he lives) and a link to his website or to one of his accounts on other social networks. Another social network, about.me, is characterized by its one-page user pro les, each with a large background image and short biography. At the moment of registration a user has to ll the suitable form with his username, email, password for the site and at a second step short biography, a short description, a pro le image and a background image.
Links to external social networks. An important feature provided by all the social networks considered in this paper is the possibility for a user to add in his pro le a link toward one of his accounts in another social site or external website. This feature is typically enabled during the creation of the user pro le. It is of particular interest in this paper because it encodes the basic information allowing the possibility of seeing di erent social sites as members of a MultipleSocial-Network environment.
Friendship. After creating a pro le, participants are asked to invite their friends to the site or to look at others' pro les and add those people to their list of friends. In Twitter, a user can follow another user, becoming his follower. Only if this user follows him back the relationship is bidirectional. Di erently from Twitter, Facebook requires approval for two people to be linked as friends. When someone links another as a friend, the recipient receives a message asking for con rmation. Indeed, Facebook friendship is bidirectional, hence, once a user accepts a friendship request of another user they become mutual friends.
Resources. A Social network resource is a Web asset such as a status update, a photo, a web link or a video created and loaded by a user in his pro le. As for LinkedIn, a user can add a resource like a new item or a new le in his pro le. He can also embed a comment, a photo, a web link or a video in a new status update. Also skills representing speci c technical expertise can be seen as a typology of resource, which are posted by users to describe their ability. This way, his connections can like it, comment it and share it on their \wall".
Actions on resources. So far, we stated that in addition to the content that members add when they create their own pro les, social network sites typically provide the possibility to share resources. After a resource is published by a user, several actions can be performed on this resource: other users can appreciate it, or re-share it, or it can be associated with a user through a mention on his pro le. Hereafter, we list the main actions a user can do on a resource according to the di erent social networks analyzed in this paper.
Once a user write a tweet in Twitter, it will appear on the homepage of all his followers, who can reply to it, make it one of their favourites or retweet it (that is, forwarding it again on their own timeline). A tweet can contain also a user mention. It can be done using the symbol @ followed by the referenced username. To categorize tweets by keyword, people use the hashtag symbol # before a relevant keyword or phrase (no spaces) in their tweets.
Clicking on like option on LinkedIn presents some di erences w.r.t. the Facebook like function. Indeed, on LinkedIn, when users click on the like link underneath the various updates, this immediately forwards that particular update out to all of the user rst level connections. The share option, instead, allows users to either redistribute the article (and partially modify it) as an update to their connections, post it to a group (or multiple groups), or forward it in a private message. Similarly to what happens in Twitter, also in LinkedIn while a user publishes a resource he can mention one of his connections with the @ symbol. He can also use a keyword as hashtag using the # symbol.
As for Flickr, by clicking on a photostream image, it is possible to open it in the interactive photopage, thus allowing users to comment it and to embed it on external websites. Moreover, images can be added to a user favourite list or to user galleries. The main Google+ page consists of a \stream" of updates, conversations and shared content. A user can make comments underneath content shared by other users, and he can appreciate contents clicking \+1" on it. Google+ provides the referencing functionality in its posts. A user can mention another user using the + or @ signs.
As for LiveJournal, users can interact with resources in di erent ways. For instance, a user can leave a comment on a post of another user or share it in his blog. He can also add to \Memories" a post. The Memories feature on LiveJournal allows the organization of favorite resources with a keyword-based archive system. Thanks to this functionality, a user can also add tags, or descriptive keywords, to his own resources.
All the features of the OSNs described in this section are mapped by our model, and this is formalized in the next section. 3
To model at an abstract level the entities described in the previous section, we use a graph. The set of nodes is partitioned into three disjoint sets P , R, and B, which correspond to the set of social pro les, the set of resources, and the set of bundles (which are resource containers), respectively.
An element of P models the pro le of a user on a social network. It consists in the tuple hurl, socialNetwork, screen-name, [personalInformation], [picture]i, where url is the Web address that identi es and localizes the prole, and socialNetwork is the commercial name of the social network which the pro le belongs to, screen-name is the name chosen by the user who registered the pro le to appear in the home-page of the pro le or when posting a resource, and, nally, personalInformation and picture are the information and the image which the user inserted as related to the pro le. The two last elements of the tuple are optional (i.e., they can be null).
The set R models resources of the Web or created by users. A resource is represented by a tuple hurl, type, [description], [date]i, where url is the Web address to access the resource, type indicates the type of the resource content, and nally, description and date, which are optional, represent the string, inserted by the who published the resource, describing the resource itself and the publishing date, respectively. For example, the most viewed video on YouTube is a resource represented as h'https://www.youtube.com/watch?v=9bZkp7q1 9f0', 'video/mp4', 'PSY - GANGNAM STYLE', '07/15/2012'i.
Our model includes the bundle set B. Indeed, commonly users do not handle a single resource, but most of the actions they do (e.g., publishing or sharing) involve more resources simultaneously. For example, a user can publish more photos or videos, can include a comment, and so on. In our model, we include all resources handled simultaneously by a user in a bundle. A bundle is represented by a tuple huri, [description], [date]i, where uri is the identi er of the bundle, description, which is optional, is the string chosen by the user to be shown with those resources and, nally, date represents the publishing date. As we will see next, we represent the inclusion of a resource into a bundle by means of containing edges.
In our model, relationships among pro les, resources and bundles are represented by direct edges of a graph. The set E of these edges is partitioned into 8 disjoint sets, named F , M , P u, S, T , Re, L, and Co.
The follow edge set F E = fps; pt j ps; pt 2 P g models the fact that in the (source) pro le ps, it has been declared a certain type of relationship towards the (target) pro le pt. This kind of edge models di erent relationships. For example, on Facebook or Flickr, it models friendships, on LinkedIn, job contacts, and, on Twitter, followers. Observe that, typically, this kind of relationship occurs between users of the same social network, because it is presumable that a social network does not have interest in promoting links to pro les of another (competitor) social network.
The me edge set M E = fps; pt j ps; pt 2 P g denotes that the user with pro le ps has declared in this pro le to have a second pro le pt. This edge allows a user to provide a link to its pro le (typically) on a di erent social network or (sometimes) on the same social network (as a sort of alias).
The publishing edge set P u E = fps; bt j ps 2 P; bt 2 Bg indicates that the user with pro le ps has published in this pro le a bundle bt. This edge models one of the typical actions a user does when enriches his/her pro le by publishing resources.
The shared edge set S E = fbs; bt j bs; bt 2 Bg speci es that the bundle bs (published by a user) is derived from an already published bundle bt. This type of edge is used when a user shares an existing bundle. Indeed, this action is represented by two edges: a publishing edge (as described before) and a shared edge from the new bundle to the existing one.
The tagging edge set T E = fps; brt; w j ps 2 P; brt 2 B [ R and w is a wordg, denotes that the user with pro le ps assigned the word w to describe a bundle or a resource br. By means of the tag mechanism, users contribute to resource labelling, which is necessary to carry out several actions on resources, such as searching or classi cation.
The referencing edge set Re E = fbs; pt j bs 2 B; pt 2 P g models the fact that a bundle bs includes a reference to the pro le pt. For example, this occurs when a tweet includes a user account name.
The like edge set L E = fps; pbrt j ps 2 P; pbrt 2 B [ R [ P g describes the information that a user with the pro le ps expressed a preference/appreciation for a bundle, a resource or another user pro le pbrt.
The containing edge set Co E = fbs; rt j bs 2 B; rt 2 Rg indicates that a bundle bs contains the resource rt. For example, when a user publishes a photo p and includes a comment c, this action is modeled by creating a bundle b with a description c, a resource p, and nally, a containing edge from b to p.
Concerning how to practically map real-life data from social networks to each component of the model, the reader can refer to [3]. In the next section, we show how this model has been exploited at application level. 4
Evaluating the accuracy of a model is a di cult task because often a golden standard misses [2]. In these cases, evaluation can be done by humans (e.g., [13, 11]) or by applying the model to an application and evaluating the results (e.g., [16]). In this section, following the latter approach, we describe how our model has been pro tably applied to two applications very relevant in the context of social network analysis.
The rst application we discuss regards the extraction of information from a multiple-social-network scenario. It is well known that any analysis activity on social network users needs a preliminary task implementing the extraction of data from social networks. In the past, several visit strategies have been adopted, such as Breadth First Search [19], Random Walk [10] or Metropolis- Hastings Random Walk [17]. In all these cases, data analysis focused on a single social network and data extraction was a quite simple task because there was not the problem of receiving data from di erent sources.
When data extraction involves di erent social networks, having a model that is able to handle indi erently data from di erent social networks is a very useful tool. In this case, it is possible to exploit a crawling task implementing the following steps. 1. Selecting the starting account (seed). This step is very important to provide data useful to the speci ed application. Usually, the starting account is randomly selected from an available pool of accounts. For particular analysis, the seed can be selected from those accounts having some characteristics, for example, being a power user (i.e., they have a number of contacts much higher than the average user [9]). 2. Building the sub-graph. In this step, the information about this account is created: it includes the user account, contacts, published resources, and so on. This step is strongly facilitated by our model. Indeed, by following the procedures described in Section ??, we map all information extracted from the di erent social networks to the components of our model (i.e., pro les, resources, bundles, and their relationships). 3. Selecting the next account. There exist several strategies to implement this step. A rst possibility is to randomly select another pro le (uniform sampling), and this is feasible whenever a social network uses an identi er for accounts and the domain of identi ers is known and limited. This occurs for example for Facebook and Twitter [7]. Another possibility consists in selecting one pro le (i.e., a node of the graph) connected with the last visited pro le by a follow edge or a me edge (see, for example, [10, 17]). Again, it is also possible to select more than one (even all) of the pro les referred above, as done for example in [4, 19]. Once one or more pro les have been selected, Steps 2 and 3 are iterated until the desired amount of data have been extracted or a stop condition has been reached.
The model de ned here has been successfully used in the SNAKE system [6], a tool supporting the extraction of data from social network accounts.
The second application that bene ted from our model concerns the problem of identifying users on the Web. A common approach to address this problem utilizes pro le matching techniques typically based on a set of identi cation properties, such as username, to nd user corresponding identity. In [5], an improvement of this approach is proposed. In particular, a new notion of pro le similarity is de ned, by combining a string similarity between the associated usernames with a contribution based on a suitable recursive notion of common-neighbor similarity. The computation of the second contribution requires to compare pro les coming from di erent social networks, which could be quite heterogeneous. The use of our model allowed us to simplify this issue and to handle all pro les in a uniform way. We can state that the success of the technique described in [5] strongly relied on the model described in this paper. 5
It is a matter of fact that the multiplicity of social networks together with users' membership overlap, result in a multiplicative e ect in terms of information power. Indeed, correlation, integration, negotiation of information coming from di erent social networks o er a lot of strategic knowledge whose bene ts are still unexplored. In this paper, we have de ned and implemented a model aimed at creating a middleware on top of existing online social networks. The goal is to provide a (conceptual) layer able to facilitate design and implementation of applications relying on the internetworking nature of online social networks. By means of two case studies, we showed the e ectiveness of the proposed model. 2. J. Brank, M. Grobelnik, and D. Mladenic. A survey of ontology evaluation techniques. In In Proceedings of the Conference on Data Mining and Data Warehouses (SiKDD 2005), 2005. 3. F. Buccafurri, G. Lax, S. Nicolazzo, and A. Nocera. A model to support design and development of multiple-social-network applications. Information Sciences, 331:99{119, 2016. 4. F. Buccafurri, G. Lax, A. Nocera, and D. Ursino. Moving from social networks to social internetworking scenarios: The crawling perspective. Information Sciences, 256:126{137, 2014. Elsevier. 5. F. Buccafurri, G. Lax, A. Nocera, and D. Ursino. Discovering missing me edges across social networks. Information Sciences, 319:18{37, 2015. 6. F. Buccafurri, G. Lax, A. Nocera, and D. Ursino. A system for extracting structural information from social network accounts. Software: Practice and Experience, 2015.
DOI: 10.1002/spe.2280. 7. M. Gjoka, M. Kurant, C. Butts, and A. Markopoulou. Walking in Facebook: A case study of unbiased sampling of OSNs. In Proc. of the International Conference on Computer Communications (INFOCOM'10), pages 1{9, San Diego, CA, USA, 2010. IEEE. 8. M. N. Jelassi, C. Largeron, and S. B. Yahia. E cient unveiling of multi-members in a social network. Journal of Systems and Software, 94:30{38, 2014. 9. S.-H. Lim, S.-W. Kim, S. Park, and J. H. Lee. Determining content power users in a blog network: an approach and its applications. Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions on, 41(5):853{862, 2011. 10. L. Lovasz. Random walks on graphs: A survey. Combinatorics, Paul Erdos is
Eighty, 2(1):1{46, 1993. 11. A. Lozano-Tello and A. Gomez-Perez. Ontometric: A method to choose the appropriate ontology. Journal of Database Management, 2(15):1{18, 2004. 12. V. S. A. Menezes, G. Zimbra~o, and J. M. Souza. Group and link analysis of multirelational scienti c social networks. Journal of Systems and Software, 86(7):1819{ 1830, 2013. 13. P. Mika. Ontologies are us: A uni ed model of social networks and semantics. In
The Semantic Web{ISWC 2005, pages 522{536. Springer, 2005. 14. D. T. Nguyen, H. Zhang, S. Das, M. T. Thai, and T. N. Dinh. Least cost in uence in multiplex social networks: Model representation and analysis. In Data Mining (ICDM), 2013 IEEE 13th International Conference on, pages 567{576. IEEE, 2013. 15. A. Papadimitriou, P. Symeonidis, and Y. Manolopoulos. Fast and accurate link prediction in social networking systems. Journal of Systems and Software, 85(9):2119{ 2132, 2012. 16. R. Porzel and R. Malaka. A task-based approach for ontology evaluation. In ECAI
Workshop on Ontology Learning and Population, Valencia, Spain, 2004. 17. D. Stutzback, R. Rejaie, N. Du eld, S. Sen, and W. Willinger. On unbiased sampling for unstructured peer-to-peer networks. In Proc. of the International Conference on Internet Measurements, pages 27{40, Rio De Janeiro, Brasil, 2006.
ACM. 18. Z. Sun, L. Han, W. Huang, X. Wang, X. Zeng, M. Wang, and H. Yan. Recommender systems based on social networks. Journal of Systems and Software, 99:109{119, 2015. 19. S. Ye, J. Lang, and F. Wu. Crawling online social graphs. In Proc. of the International Asia-Paci c Web Conference (APWeb'10), pages 236{242, Busan, Korea, 2010. IEEE.