Overview and Analysis of Personal and Social Tagging Context to construct User Models Karin Schoefegger Knowledge Management Institute, Graz University of Technology Inffeldgasse 21a, 8010 Graz, Austria k.schoefegger@gmail.com Michael Granitzer Knowledge Management Institute, Graz University of Technology Know-Center Inffeldgasse 21a, 8010 Graz, Austria mgrani@know-center.at ABSTRACT The quality and user acceptance of personalized services such as personalized information retrieval and navigation or con- tent recommendation depends depends besides the personal- ization mechanism on the quality, validity and accuracy of the employed user model. In literature a variety of user model construction methods based on tagging activity in social tag- ging systems (STS) are presented, relying on different user contexts, e.g., the personal or social context. But up to now there is neither a concise overview of existing construction methods available nor a deeper analysis and discussion of the differences between these models. Such an analysis would for example ease evaluation but also enable system design- ers to choose the most appropriate one. Our work tackles this problem by providing a short overview of state-of-the art user model construction methods which employ social tags. This is followed by a statistical comparison of four different user model construction methods for STS based on tag-frequency. This analysis unveils that depending on the method chosen (based user’s personal tagging behavior as well as community- based social strategies), the user model consists of different tags and tag frequency rankings, thus services employing dif- ferent models will lead to different results. Author Keywords User Modeling, Social Tagging, Recommender Systems, Per- sonalized Information Retrieval ACM Classification Keywords H.1.2 User/Machine Systems: Human Factors H.1.2 Infor- mation Systems: Models and Principles—Human informa- tion processing General Terms Algorithms, Human Factors. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. CaRR 2012, February 14, 2012, Lisbon, Portugal. Copyright 2012 ACM 978-1-4503-1192-2/12/02...$10.00. INTRODUCTION In the last years the web has witnessed an explosion of in- formation created and shared, by individuals and through so- cial interaction, such as for example in social tagging systems or social streams. This amount of information has gener- ated a huge need for more effective access to the information, especially since each user has different expectations, goals, knowledge, information needs and desires to be satisfied. One way to ease the access is by personalization. To enable a sys- tem to adapt to e.g. the user’s interests, the system usually builds a user model, an accurate machine-readable represen- tation of the user. In social tagging systems, which enable a user to annotate a resource with a freely chosen keyword (tag) for later reuse and sharing, the user’s interests are commonly modeled by using the tags available in the system. In literature, a variety of user model construction methods have been presented but to the best of our knowledge, there is yet no concise overview of common techniques available. E.g. several methods build the user model solely based on the personal tagging context (only the tags the person used) or the social context (the tags all users used for a specific user’s resources). Thus, in the first part of this work we present a literature overview and based on that, we present four typical user model construction methods and variations of it which can be used to enhance personalization in a social tagging system. Each of these models can be used for example to personalize resource ranking for retrieval with the FolkRank algorithm [13]. FolkRank provides the possibility to adapt the preference vector for a random surfer component to express user preferences by giving a higher weight to components that represents the user’s preferences. It is expected that different user models (based on their context) lead to differently ranked results. This raises the question whether typical UM tech- niques lead to similar user model representations or whether they exhibit a different user model depending on what tech- nique is chosen which in turn is expected to lead to differ- ent search result rankings. Thus, in a second part, we present and discuss a statistical analysis of the comparison of the four chosen user model construction methods. We thereby focus on general statistics such as the ’tag richness’ of a user model, the similarity between different possible models for one user as well as the correlation between the similarity of the possi- ble user models for one user and the resource sharedness of this user.

Contribution of this work is the following: • We provide a short overview of user model construction methods in Social Tagging Systems. • We present a set of four possible and popular user model- ing strategies to personalize services in STS. Two models are based solely on the user’s tagging behavior, two repre- sent the user model based on the tagging behavior of the community (userbase of the system). • We show that the more tags are assumed to be part of a tag-based user model, the more specific a user model be- comes and the different the four possible user models be- come. For personalized information retrieval a rich, spe- cific user model is preferred, thus it plays an important role which user model is chosen to be fed into the personaliza- tion mechanism. • We furthermore show that the more resources a user shares with the community, the more the personal and community- based profiles differ. For these users, it needs to be further investigated which user model is the most useful one, de- pending on the use case (e.g. personalized IR, expertise search, navigation). SOCIAL TAGGING Tagging has gained a major success in the web 2.0 as it plays an important role of helping user manage their resources. Users are encouraged to add tags to describe a website, a publica- tion, a music track, a picture, etc. and to share these resources tags with other people. These tags indirectly reflect a user’s interests, concerned topics, activities in daily life, and many more. Thus social tagging activities can serve as as an inter- esting source of information to build a user representation for any kind of personalized service. Definitions In the following we describe a few definitions of terms that are used throughout this work. • A User Library ULib(u) is the set of all resources a user u has annotated within a social tagging system. • An Author Library ALib(u) is the set of resources who are authored by a specific user u. E.g. an academic publi- cation annotated with ’myown’ in Bibsonomy 1 or added to the folder ’My Publications’ in Mendeley2. Usually, ALib(u) ⊂ ULib(u) and one resource r can have more than one author. • A Folksonomy F (R U T TAS) is the central data struc- ture of a social tagging system like Bibsonomy (academic publications as well as web resources), which is commonly seen as a lightweight classification structure built from so called tag annotations (TAS) added by different users to their resources. A folksonomy consists thus of a set of users U, a set of tags (i.e. freely chosen keywords) T , and a set of resources R with ULib(u) ⊆ R∀u ∈ U, together with a ternary relation TAS ⊆ U × T × R between them. 1http://www.bibsonomy.org 2http://www.mendeley.com • A Personomy P (u) = F (ULib(u) u T (u) TAS),u ∈ U is a subset of F (R U T TAS), built only from tag annotations of a sin- gle user (person). • The Tag Sharedness TS(t) of a tag t ∈ T is given by |{u ∈ U|t ∈ T (u)}|, thus is the amount of users who used a specific tag t . If we limit the tags of F (R U T ) to F2(R U T2] where T2 = {t ∈ T|TS(t) ≥ 2} is the set of tags with tag sharedness greater than 1, this eliminates on the one hand e.g. typical problems of a tag vocabulary such as misspellings etc. but also on the other hand increase the ’information value’ of a tag in a user model. • The Resource sharedness RS(r) of a resource r ∈ R is given by |{u ∈ U|r ∈ ULib(u)}|, thus is the amount of users who have a specific resource r in their library. OVERVIEW OF USER MODELING APPROACHES FOR SO- CIAL TAGGING SYSTEMS In social tagging systems, it is generally assumed that anno- tating a resource is a good indicator for the current interests of a user. E.g. if a large number of a user’s tagging activi- ties include the tag ’sports’, the user is likely to be interested in sports-related content. Some work also models the user’s expertise- or knowledge-based on the user’s tagging behavior ([14], [28], [4]). A categorization of the user model construction is usually somehow problematic as some approaches apply more than one construction technique. For example, in concept-based methods often clustering is applied to identify clusters repre- senting one concept. Also, this is not a complete overview but we rather aim at presenting the most common methods together with some prime examples. Tag-Frequency-based User Modeling In most approaches presented in literature, a tag-based user model representing the user’s interests or expertise is usually provided in form of a weighted tag vector. In its simplest form, a weight is given by the frequency of the tag in the user’s personomy. Tag-frequency-based user models are con- structed in different ways, the following two main approaches can be distinguished: Firstly, the user model is based on tags extracted from a user’s personomy, thus on the tags the user has directly assigned to annotate resources. The work of [18] or [3] follow the naive approach which simply represents the user in form of a tag (frequency) vector which indicates that user u has used a tag t (a certain number of times) to annotate an item. The personomy-based user model depends on the fact that the user has to collect a sufficient amount of annotation data such that the system can infer a useful user model. [1] present a more lightweight approach which builds the user model based upon the tags that other users have added to the resource the specific user clicks on. Similarly, in the work of [9] or [12] a personalization strategy for IR based on folksonomy data is presented, the user model is enriched with the tags other users