Social tagging represents an innovative and powerful mechanism introduced by social Web: it shifts the task of classifying resources from a reduced set of knowledge engineers to the wide set of Web users. Tags generate folksonomies; in the current popular social tagging systems (such as delicious or Bibsonomy), they are di cult to manage, modify, and visualize in dynamic and personalized ways. The aim of this paper is to describe Folkview, an innovative way to conceive a folksonomy in terms of a multi-agent system. Folkview is able to support speci c modular tools for personalizing customized and dynamic visualization features allowing users to simply update, manage and modify a folksonomy.
Social tagging systems are characterized by the active participation and interaction of users, which upload, share and freely annotate with labels, known as tags, a huge amount of resources, explicitly inducing on them personal classi cations. Although these systems are widely used and personal annotations represent a democratic, powerful and easy way of classifying resources, they su er from different issues:
The creation of personalized views, which may display a limited, well de ned and personalized sub-portion of an entire hyperspace is something that has already been considered in di erent settings. To implement this strategy, most of traditional Web browsers become to o er personalized views, so-called start pages such as, e.g. NetVibes1, My Yahoo2 and iGoogle3. Some extensions to
these examples are adaptive bookmarking systems such as PowerBookmarks [
These applications highlight that a crucial task for the developers of nowadays Web application is how to model and create speci c tools for providing personalized views to the users. All the types of social tagging systems should deal with these compelling and open challenges, expanding their capabilities and enhancing possible multiple visualizations, in order to achieve (a) a more e ective comprehension of the semantic relations of a folksonomy, (b) a more useful navigation through the involved elements, and (c) the manipulation of the existing relations among tags and resources according to the user needs.
The folksonomies (and the personomies) are generally visualized as a tag
cloud: in spite of this, the work [
The user navigates her tag cloud (shown on the left); when she selects a speci c tag (\Webdesign", in our case), the number of bookmarks related to the chosen tag and the list of resources annotated with it are shown. The navigation may continue by clicking on each resource, tag or user, but { the tag cloud is not adaptive; { personalized views cannot be created; { it is not possible to simply modify the personomy, or the personal view of the folksonomy (for example, renaming a tag for a set of resource or merging two or more tags on a unique label).
These limitations are partially ascribable to the static nature attributed in
literature to a folksonomy; in fact, it has been de ned in terms of nite sets
of users, resources and tags [
The main aim of this work is to propose a novel, distributed, modular system called Folkview, whereby a folksonomy is conceived dynamically through the use of multiple agents. These agents will be capable of { managing the structural and semantic properties; { cooperating for obtaining common objectives; { o ering personalized and dynamic views.
The paper is organized as follows: in next Section 2 we discuss related work, in Section 3 we present the formal, multi agent-based model at the basis of Folkview, while, in Section 4, we discuss its dynamic features with respect to authoring and personalized views. Finally, nal considerations end the paper. 2
Early de nitions of folksonomy [9{11] are related to the user activity of
annotating resources with metadata for her own individual aims, and/or for sharing
them in a community. In these de nitions, only three kind of entities (users,
resources and tags) and the relations among them, called tas (tag assignments),
are considered, instead of any dynamic aspect of visualization and manipulation.
An extended de nition of the previous ones is given in [
As observed in the introduction paragraph, a folksonomy is usually represented
by a tripartite graph or network, but this leads to another issue related to the
complexity of the nature of the graph itself. Various researches have dealt with
this problem, projecting a folksonomy on simpli ed structures. For example,
in [
A few research projects have addressed some of them: in [
Another project is TagGraph4, a folksonomy navigator which visualizes the relationships between Flickr tags. User may enter a Flickr username or a tag, and the graph sets out drawing itself automatically; after this early step, she may navigate through related tags or among related images, but could not manipulate her personomy.
The mentioned projects are by all means interesting attempts of interactive visualizations of folksonomies; nevertheless they do not provide neither personalized views nor e ective dynamic changes according to the user needs or preferences. 3
Traditionally, given the sets U , T and R respectively of users, tags and resources, a folksonomy is de ned as the set of tag assignments (tags, for short) (ui; rj ; tk) 2 U T R, where i = 1; : : : ; jU j; j = 1; : : : ; jT j; k = 1; : : : ; jRj, each of them indicating that user ui has tagged the resource rj with tk. User pro les, functions, metrics or semantic relations among users, tags, resources and tas are not intrinsic properties of the folksonomy, but may be (or not) applied by the system which hosts the folksonomy. We indicate this traditional concept of folksonomy as static folksonomy F . In order to de ne a F , we identify three classes of sets: { Tui;rj 2 T is the set of tags used by ui on rj ; { Rui;tk 2 R is the set of resources tagged by ui with tk; { Utk;rj is the set of users that tagged rj with tk.
Each set represents a structural component of the folksonomy, and we call it structural ; the tags are grouped associating to them a semantic label for identifying their meaning in that dimension. A graphical example of 6 sets of tags is given in Figure 2, on the left.
The rst three linear paths contain the resources tagged by user u1, using respectively t1, t2 and t3. So, the labels associated with them are respectively u1; t1, u1; t2 and u1; t3.
De nition 1. A structural dimension is a labeled path
Dui;rj = (V; E; ) where V = Tui;rj is the set of vertices, E is the set of edges, (e) = (ui; rj ) 8e 2 E is an edge labeling, and degree(tk) = 0; 1; 2 8tk 2 Tui;rj . In particular, degree(tk) = 0; 1; 2 only if jTui;rj j = 1.
{ supporting tag suggestions, or recommendations; { inferring knowledge about the user pro le, her habits, preferences, and skills; { for identifying similar users, resources or tags.
In summary, folksonomies add semantics on the data. For this reason, we propose a new concept of folksonomy, conceived as a dynamic entity, organized as an universe of inherently autonomous computational sub-entities, which interact with each other by sending messages and reacting to external stimuli by executing some prede ned procedural skills. Various authors have proposed di erent de nitions of agents. In our setting an agent is formally de ned as follows. De nition 3. An agent A = (Ts, En, Re, Ac) is a quadruple where { T s represents its topological structure; { En = f 1; 2; : : :g de nes its local environment; { Re = f 1; 2; : : :g is the nite set of incoming requests; { Ac = f 1; 2; : : :g is the discrete, nite set of possible actions. Ts and En represent the passive part of the agent, while Re and Ac its active part.
Finally, we can introduce the de nition of the dimension Dui;rj based on the structural dimension Dui;rj .
De nition 4. A dimension D = (T s; En; Re; Ac) is an agent where { T s = Dui;rj ; { En = fui; rj ; t1; : : : ; tng; { Re = f;g, initially; { Ac = fadd-tag; delete-tag; modif y-tag; : : :g Analogously, we can de ne new classes of agent dimensions, not only for structural dimensions. New dimensions can be created directly from the user, or computed by the system applying speci c metrics, or generated applying ontological models; each dimension can contain other dimensions; each dimension associates a semantics to the set of grouped entities.
De nition 5. A folksonomy F is a multi-agent system formally described as a labeled multigraph of agent entities, organized in semantic contexts, called dimensions.
n F = [ i=1
D i All in a folksonomy is a computational agent, equipped with a set of local variables, that de ne its internal state, and a modular and extensible set of procedural skills. So, for example, each user is represented in a folksonomy by an user agent: it knows the resources tagged by the user, and the used tags; but it also contains and manages the user pro le, and it is able to calculate speci c local metrics for her, such as the average number of tags applied on a single resource, the average time spent on a resource, the tagging date, etc. They can further communicate with the other agents present in the personomy and in the folksonomy, such as the tag agents, or the resource agents, or the same dimension agents.
Some semantic connections may be inferred by F , applying opportune metrics, for example a similarity function for identifying neighbors tags, users or resources, or ontological relations. 4
It is simple recognize in the labeled multigraph contained in the de nition of
F the zz-structures [
The peculiarity of such structures derives from the relation among their component elements: data is stored into cells, that may contain very di erent type of contents, which are connected with links of the same color into linear sequences called dimensions. A single series of connected cell among one dimension is called rank, while the starting and the ending cells of a rank are called headcell and tailcell. There is also a restriction according to which for any dimension, each cell can connect almost two other cells following the direction of the dimension.
As discussed in literature [
In [
In Figure 3 (left) we show a H-view, on two dimensions, Du2;t4 and Du2;t1 , extracted from the folksonomy shown in Figure 2. We note the presence of a black triangle symbol, in two positions, corresponding to selected/not selected: these triangles are associated to scripts related to the session agent of the current visualization, and represent the mean to interact with the cell-agent. When selected, the session agent asks to the chosen resource (r9, in our example) the set of actions Ac that can be activated on it. In order to satisfy this request, r9 sends a multicast message to all the dimensions in which it is included, and a run-time created contextual menu, organized in three meta-categories (views, metrics and semantics) is shown. The rst category is concerning the di erent kind of possible views while the other two categories of functions o ered by the menu, are related to the computation of an extensible set of metrics, and to the application of opportune semantic relations and ontologies in order to generate, for example, speci c recommendations on content, tag and user.
In our example, the user selects the menu item views and then, from its sub-menu, the menu item m-extended star view (where m = 3). The related 5extended star view is displayed in Figure 3 (right): we can note that the cell r9 is connected to the following: { three labeled edges related to the dimensions Du1;t1 , Du2;t1 Du2;t4 ; { one labeled edge related to the dimension Dui;rj , i.e. the tags (t4 and t1) which other users applied on the same resource r9; { one labeled edge concerning the dimension Dtk;rj , i.e. the users (u1 and u2) which annotated the resource r9 with the same tag.
Comparing the two visualizations, it is clear that the 3-extended star view provides a deep insight of all the dimensions connected to a given focus cell.
Other features, not displayed in Figure 3, regard the possibility to dynamically change, at local or global level, the features of each agent, simply clicking directly on the visualized item and applying modi cations. To this extent we can highlight that due to the agent-based technology the folksonomy grows and changes according to the user contributes, and can be shared with the other users. 5
In this paper we have proposed an innovative way to conceive a folksonomy in terms of a multi-agent system, rst de ning a formal model and then showing Folkview. Such system can be used to simply display personalized user views, to create personalized and adaptive paths for users and to modify the associations between tags and resources.
Up to now we have built a partial, but modular and extensible, prototype, based on a public dataset taken from delicious, and that implements the structural aspects of the considered folksonomy, adding main existing metrics functions, and using both server-side and client technology.
As future work we want to extend the prototype to all the main functionality we discussed, focusing our attention on a semantic personalization. In particular, we plan to make user tests to assess the impact and the e ectiveness of the proposed tool, comparing particular user-tasks on our proposal system than what already exists as social tagging system. Furthermore, although we started with a speci c dataset, we intend to extend our tool in order to extract data from a large number of social tagging systems.