How can we make use of the personal information a single user is spreading all over the Social Web every day? In this paper we investigate what is needed from a user model point of view to support user data sharing and aggregation to enhance personalization and recommendation services. We present a study of 17 social applications to de ne requirements and attributes for a common user model that allows sharing of user data and analyze what is needed to enhance user model aggregation approaches. As a result, we present a comprehensive user model especially tted to the needs of the Social Web. Furthermore, we present a WordNet for the user modeling domain as part of the user model to support user model aggregation.
Every day, people in the Social Web create 1.5 billion pieces of information
on Facebook, over 140 million tweets on Twitter, upload more than 2 million
videos on YouTube and around 5 million of images to Flickr3. This huge amount
of social data attracts researchers who want to use it to learn more about user
preferences and interests, and enhance recommendation and personalization
systems. What most current system have in common is that they use data from a
single application and depend on su cient user information (user behavior or
ratings) to produce good results [
In this paper, we want to give answer to those questions by analyzing user models from di erent Social Web applications and draw conclusions about the diversity and type of user information that such a generic user model should have. We therefore discuss existing work and motivate a semantic Social Web User Model (SWUM). Requirements and structure of SWUM will be introduced in Section 3 and is based on the extensive analysis of 17 Social Web applications in Section 3.1 and 3.2. In Section 3.3 we also carefully investigate what is needed to enable an easy, automated, aggregation process. To give a better understanding of the intended use of the SWUM we present a use case in Section 4. The main contributions of this paper are an extensive analysis of requirements of today's Social Web applications regarding stored user data and the introduction of a new Social Web user model that is: { generally adapted to the needs of Social Web applications and { that allows an easy data sharing between applications. 2
Until the turn of the millennium, most personalization and recommendation
research focused on user information available in one application and how to use
this information to enhance personalization quality. With the in uence of the
Social Web, or Web 2.0, and the fact that user information is highly distributed
over several applications, research started to explore cross-system personalization
approaches. This research can be roughly classi ed into two major directions [
The work presented in this paper is in alignment with the rst direction, the
centralized approach. This approach can also be subdivided into two
aggregation strategies. The rst strategy proposes the use of standardized user models
which all involved applications must agree on. The second strategy deals with
the mediation of di erent user model representations using meta-models that
connect user data from one application with data from another application, in
the same domain, or across domains. The standardization approach involves no
computational e ort to aggregate data as all data already is in the same format.
An e ort in this direction is the General User Modeling Ontology (GUMO)
created by Heckman et al. [
The second strategy is to build meta-models that allow de ning how
applicationdependent user data corresponds to user data from another application. This has
the advantage that applications are not forced to adopt a prede ned generic user
model and can rely on their own model. In [
To summarize: both strategies and the presented related work have shortcomings. Because of the big di erences, regarding the covered user information and representation forms in di erent applications, the development of a commonly accepted ontology, covering all aspects of user modeling for all domains seems not feasible. The meta-model approach, without automatic aggregation mechanisms, is only applicable in small settings where only a few applications are connected and not for the Social Web. We therefore propose a middle way: We need a new `common' user model that combines aspects of the presented approaches and focuses on a special domain, the Social Web. Also, the user model should support automatic aggregation by de ning a structure that allows nding relations between di erent user model concepts and allows for a exible extension of the model. 3
To de ne a user model for the domain of the Social Web, we rst have to understand the demands of social web applications on user models. Therefore, we did an extensive survey of the modeled user information of 17 well-known Social Web applications. The list of analyzed applications is shown in Table 1. The applications were chosen because of their size and level of awareness (number of users, global distribution). To be able to consider local di erences, we also included applications that are strong in only one or two regions (Orkut in South America, Lokalisten and StudiVZ in Germany). We also selected Social Web applications from di erent kinds of domains, photo- and video-sharing platforms, short-message services, social networks, etc. To decide if the user information stored by an application is of importance, we picked at least two Social Web applications from the same domain. Facebook http://www.facebook.com Myspace http://www.myspace.com Windows Live http://home.live.com YouTube http://www.youtube.com Flickr http://www.flickr.com Yahoo http://de.yahoo.com Picasa Web http://picasa.google.com StudiVZ http://www.studivz.net Digg http://www.digg.com Yelp http://www.yelp.com Lokalisten http://www.lokalisten.de Orkut http://orkut.com Identi.ca http://identi.ca LinkedIn http://www.linkedin.com Vimeo http://www.vimeo.com Xing http://www.xing.com LastFM http://www.last.fm
For each evaluated application, we collected the type of information and the internal attribute name. Table 2 shows the type of user information and where the information was found on the Web page. The internal attribute names, used by each application are particularly important as they are later used to de ne and name the attributes of the Social Web User Model (SWUM).
To be able to create our SWUM, we rst have to decide which type of information, which user model dimensions, should be part of the model and which attributes in the di erent dimensions should be supported. 3.1
After collecting all the information, the rst step is to determine the user model
dimensions that our user model has to cover. As shown in GUMO, a lot of
dimensions exist, but not all of them are required in the context of the Social Web.
Several dimension are mentioned and discussed in the literature. We present a
consolidated taxonomy that bases on [
Based on this user taxonomy, we checked all 17 applications if they cover
these dimensions. Fig. 1 shows that social applications only cover some
dimensions. All of the applications maintain Personal Characteristics and most of
them also use Interests and Preferences information. Not used at all are the
dimensions Individual Traits and Mental and Physical State which are more used
in educational systems than in Social Web applications [
The usage of Knowledge and Background and Context depends on the focus
of the social application. Social business applications, like LinkedIn or Xing,
support the Knowledge and Background dimension as users can enter their college
degree, areas of profession, etc. The support for the dimension User Behavior
is not easy to work out, as user behavior usually is an implicit feature and not
displayed on the user pro le page of an application. It can be assumed, though,
that almost all applications track user behavior on their site. A positive
exception is \Google Dashboard" 4 where a user gets an easy overview of the stored
personal information e.g. previous search behavior. The User Behavior
dimension, although it is an important piece of adaptation and personalization, is to
complex to be part of a generic Social Web User Model. For this purpose we
recommend a specialized approach with an extra user behavior ontology as
presented in [
From this analysis it follows that a main requirement for Social Web user model is, that it has to cover the user dimensions Personal Characteristics, Interests, Knowledge and Behavior, Needs and Goals and Context (Social Context, Location). Accordingly, these dimensions are part of our SWUM. 3.2
After we selected the dimensions to be covered, we have to de ne the attributes that the user model should support.
The procedure for the attribute selection is similar to the procedure used to select the dimensions. We checked the di erent attributes of the di erent applications. Fig. 2 gives an example for thePersonal Characteristic dimension. It shows an excerpt of the attributes and how often they occur in the analyzed social applications. In this way, we selected a set of attributes for each dimension. An example for the Personal Characteristic dimension is shown in Fig. 3. The Personal Characteristic is divided into two main concepts namely Demographics and Contact Information. The concept Location is a helper concept to model locations and link certain information, e.g. places lived, to it.
Contact'Information' • First&name:&string& • Middle&name:&string& • Last&name:&string& • Full&name:&string& • Nickname:&string& • Username:&string& • Maiden&name:&string& • Living&in:&List&of&Locations& • Places&lived:&List&of&Locations& • Current&City:&Location& • Hometown:&Location& • Work&Phone:&int& • Home&Phone:&int& • Mobile&Phone:&int& • Home&Fax:&int& • Work&Fax:&int& • Personal&Email:&string& • Work&Email:&string& • Personal&Homepage:&string& • Work&Homepage:&string& • IM:&string& &
Demographics' • Gender:&string& o Female:&bool& o Male:&bool& • Birthday:&date& o Day:&&int& o Month:&int& o Year:&int& • Birthplace:&Location& • Language:&string& • Other&Languages:&string& • Family&status:&string& • Education:&Education& • Employment:&Employment& • Employment&History:&List&of&Employments& Location' • Country:&string& • State:&string& • City:&string& • Street:&string& • House&number:&int& • Postal&code:&int& An important outcome of the attribute distribution analysis was that often similar information is stored by most applications, but in di erently named attributes, e.g. name (Yahoo) and real name (LastFM) or homepage (LastFM) and website (Flickr). This problem of attribute name heterogeneity complicates a possible aggregation using a Meta-Model strategy. To cover that problem, we decided to extend our model with a WordNet like lexicon called User Model Word Net (UMWN). WordNet de nes word sense relations between words. If a word represents a user attribute, the relatedness between di erent attributes can be acquired easily. However, many user attributes are not de ned in WordNet. Moreover, many terms in WordNet are useless for user pro le aggregation. Hence, the standard WordNet does not help, thus, we designed a reduced WordNet, specialized to serve the user pro le aggregation and initially based on the attribute distribution of our analysis. The decision to use a WordNet based structure comes from the fact, that WordNet has a exible and well-de ned lexicon schema, which is publicly known and accepted. The user model terms can be linked to each other accurately by using the properties de ned in WordNet. An example is depicted in Fig. 4 where the word sense relations for name and date are shown.
The UMWN is an important step for an automatized aggregation of di erent user models. It de nes di erent types of word relations. The \Name" concept describes the relations between di erent types of name attributes that can occur in a user model. The concept \full name" consists of di erent subclasses like \ rst name", which has several synonyms (\given name" or \forename"). UMWN is stored in RDF(s)/OWL. Using ontology structures has the advantage that such a model is not static and can be easily extended. Our UMWN is extensible, towards not only to the individuals, but also to the schema of UMWN. Because of the highly distributed and heterogeneous user information in di erent user models, extensibility is an important feature. The UMWN contains currently ca. 520 syn sets where around 200 are unique in the User Model WordNet and not part of the common WordNet. It also contains over 100 antonyms and homonyms and 200 meronyms. 4
Use Case: Pro le Aggregation with the SWUM To outline the intended usage and functionality of the SWUM (which includes the UMWN) we want to exemplary explain the steps needed to aggregate a Facebook user model and a LastFM user model. The aggregation is a two-step process which we want to explain by the example of the website/homepage attribute shown in Fig. 5. First step is to connect the LastFM attributes to the SWUM (see Fig. 5a). The LastFM user model has the attribute \homepage" which can be directly linked to the SWUM, with a concept match of 100%. The Facebook pro le (Fig. 5b) contains the attribute \website" which is also part of our SWUM and thus, the attribute can also be linked to the SWUM without any extra e ort.
a) Aggregation of the LastFM profile with the UMWN b) Aggregation of the Facebook profile with the UMWN
The second step is then to directly connect the LastFM and Facebook user model as shown in 6. Based on the previously shown aggregation, connecting both models is straightforward. Revisiting the homepage/website example, these attributes can be directly linked because of the UMWN. The UMWN de nes a synonym relation between the concepts \homepage" and \website", thus the LastFM and Facebook attribute can be directly linked with a match of 100%.
The aggregation of attributes that are not part of the SWUM can be done not only using the attribute name but also using the attribute content. So could an analysis show that the LastFM attribute \real name" often contains the users' full name and thus a connection with the SWUM/UMWN attribute \full name" can be done. Or the missing attributes can be added to the SWUM which is easy to do as it is a exible RDF/OWL structure. 5
In this paper we wanted to answer the question what are the requirements of the Social Web for a user model to pro t from the available distributed user information. We present a new user model, the Social Web User Model (SWUM) that is tted to the needs of the Social Web. We therefore conducted an extensive analysis of 17 social applications and to specify requirements, which dimensions and attributes are needed, for a Social Web user model. Based on this analysis we de ned the dimensions a Social Web user model must cover and explained how the decision process was conducted. The analysis showed, that a Social Web user model only needs to cover certain dimensions of the user, namely Personal Characteristics, Interests, Knowledge and Behavior, Needs and Goals and Context (Social Context, Location). We also presented the procedure to de ne the attributes of such a Social Web user model. To cover the problem of attribute heterogeneity throughout di erent social applications, we also equipped our model with a reduced WordNet that is especially tailored to the area of user modeling, the User Model Word Net (UMWN). The complete SWUM and UMWN model is based on RDF/OWL and thus easy to extend and reuse.