Information Retrieval in Folksonomies: Search and Ranking

Andreas Hotho,1 Robert J schke,1,2 Christoph Schmitz,1 Gerd Stumme1,2 a

1

Knowledge & Data Engineering Group, Department of Mathematics and Computer Science, University of Kassel, Wilhelmsh her Allee 73, D34121 Kassel, Germany o http://www.kde.cs.uni-kassel.de 2 Research Center L3S, Expo Plaza 1, D30539 Hannover, Germany http://www.l3s.de

Abstract. Social bookmark tools are rapidly emerging on the Web. In such systems users are setting up lightweight conceptual structures called folksonomies. The reason for their immediate success is the fact that no specic skills are needed for participating. At the moment, however, the information retrieval support is limited. We present a formal model and a new search algorithm for folksonomies, called FolkRank, that exploits the structure of the folksonomy. The proposed algorithm is also applied to nd communities within the folksonomy and is used to structure search results. All ndings are demonstrated on a large scale dataset.

1 Introduction
Complementing the Semantic Web effort, a new breed of so-called Web 2.0 applications is currently emerging on the Web. These include user-centric publishing and knowledge management platforms like Wikis, Blogs, and social resource sharing tools. These tools, such as Flickr3 or del.icio.us,4, have acquired large numbers of users within less than two years.5 The reason for their immediate success is the fact that no specic skills are needed for participating, and that these tools yield immediate benet for each individual user (e.g. organizing ones bookmarks in a browser-independent, persistent fashion) without too much overhead. Large numbers of users have created huge amounts of information within a very short period of time. The frequent use of these systems shows clearly that web- and folksonomy-based approaches are able to overcome the knowledge acquisition bottleneck, which was a serious handicap for many knowledgebased systems in the past. Social resource sharing systems all use the same kind of lightweight knowledge representation, called folksonomy. The word folksonomy is a blend of the words taxonomy and folk, and stands for conceptual structures created by the people. Folksonomies are thus a bottom-up complement to more formalized Semantic Web technologies, as they rely on emergent semantics [11, 12] which result from the converging use of the same vocabulary. The main difference to classical ontology engineering approaches is their aim to respect to the largest possible extent the request of non-expert users not
http://www.ickr.com/ 4 http://del.icio.us 5 From discussions on the del.icio.us mailing list, one can approximate the number of users on del.icio.us to be more than three hundred thousand.
3

to be bothered with any formal modeling overhead. Intelligent techniques may well be inside the system, but should be hidden from the user. A rst step to searching folksonomy based systems complementing the browsing interface usually provided as of today is to employ standard techniques used in information retrieval or, more recently, in web search engines. Since users are used to web search engines, they likely will accept a similar interface for search in folksonomy-based systems. The research question is how to provide suitable ranking mechanisms, similar to those based on the web graph structure, but now exploiting the structure of folksonomies instead. To this end, we propose a formal model for folksonomies, and present a new algorithm, called FolkRank, that takes into account the folksonomy structure for ranking search requests in folksonomy based systems. The algorithm will be used for two purposes: determining an overall ranking, and specic topic-related rankings. This paper is organized as follows. Section 2 reviews recent developments in the area of social bookmark systems, and presents a formal model. Section 3 recalls the basics of the PageRank algorithm, describes our adaptation to folksonomies, and discusses experimental results. These results indicate the need for a more sophisticated algorithm for topic-specic search. Such an algorithm, FolkRank, is presented in Section 4. This section includes also an empirical evaluation, as well as a discussion of its use for generating personal recommendations in folksonomies. Section 5 concludes the paper with a discussion of further research topics on the intersection between folksonomies and ontologies.

2 Social Resource Sharing and Folksonomies

Social resource sharing systems are web-based systems that allow users to upload their resources, and to label them with arbitrary words, so-called tags. The systems can be distinguished according to what kind of resources are supported. Flickr, for instance, allows the sharing of photos, del.icio.us the sharing of bookmarks, CiteULike6 and Connotea7 the sharing of bibliographic references, and 43Things8 even the sharing of goals in private life. Our own system, BibSonomy,9 allows to share simultaneously bookmarks and bibtex entries (see Fig. 1). In their core, these systems are all very similar. Once a user is logged in, he can add a resource to the system, and assign arbitrary tags to it. The collection of all his assignments is his personomy, the collection of all personomies constitutes the folksonomy. The user can explore his personomy, as well as the personomies of the other users, in all dimensions: for a given user one can see all resources he had uploaded, together with the tags he had assigned to them (see Fig. 1); when clicking on a resource one sees which other users have uploaded this resource and how they tagged it; and when clicking on a tag one sees who assigned it to which resources. The systems allow for additional functionality. For instance, one can copy a resource from another user, and label it with ones own tags. Overall, these systems provide a very intuitive navigation through the data. However, the resources that are displayed are usually ordered by date, i. e., the resources entered last show up at the top. A more sophisticated notion of relevance which could be used for ranking is still missing.
6 9

http://www.citeulike.org/ http://www.bibsonomy.org

http://www.connotea.org/

http://www.43things.com/

Fig. 1. Bibsonomy displays bookmarks and BibTeX based bibliographic references simultaneously.

2.1 State of the Art There are currently virtually no scientic publications about folksonomy-based web collaboration systems. The main discussion on folksonomies and related topics is currently taking place on mailing lists only, e.g. [3]. Among the rare exceptions are [5] and [8] who provide good overviews of social bookmarking tools with special emphasis on folksonomies, and [9] who discusses strengths and limitations of folksonomies. In [10], Mika denes a model of semantic-social networks for extracting lightweight ontologies from del.icio.us. Besides calculating measures like the clustering coefcient, (local) betweenness centrality or the network constraint on the extracted one-mode network, Mika uses co-occurence techniques for clustering the folksonomy. There are several systems working on top of del.icio.us to explore the underlying folksonomy. CollaborativeRank10 provides ranked search results on top of del.icio.us bookmarks. The ranking takes into account how early someone bookmarked an URL and how many people followed him or her. Other systems show popular sites (Populicious11) or focus on graphical representations (Cloudalicious12, Grafolicious13 ) of statistics about del.icio.us. Confoto,14 the winner of the 2005 Semantic Web Challenge, is a service to annotate and browse conference photos and offers besides rich semantics also tagging facilities for annotation. Due to the representation of this rich metadata in RDF it has limitations in both size and performance. Ranking techniques have also been applied in traditional ontology engineering. The tool Ontocopi [1] performs what is called Ontology Network Analysis for initially populating an organizational memory. Several network analysis methods are applied to an already populated ontology to extract important objects. In particular, a PageRank-like [2] algorithm is used to nd communities of practice within sets of individuals represented in the ontology. The algorithm used in Ontocopi to nd nodes related to an individual removes the respective individual from the graph and measures the difference of the resulting Perron eigenvectors of the adjacency matrices as the inuence of that individual.
10 13 11 http://collabrank.org/ http://populicio.us/ http://www.neuroticweb.com/recursos/del.icio.us-graphs/ 14

http://cloudalicio.us/ http://www.confoto.org/

12

This approach differs insofar from our proposed method, as it tracks which nodes benet from the removal of the invidual, instead of actually preferring the individual and measuring which related nodes are more inuenced than others. 2.2 A Formal Model for Folksonomies A folksonomy describes the users, resources, and tags, and the user-based assignment of tags to resources. We present here a formal denition of folksonomies, which is also underlying our BibSonomy system. Denition 1. A folksonomy is a tuple F := (U, T, R, Y, ) where U , T , and R are nite sets, whose elements are called users, tags and resources, resp., Y is a ternary relation between them, i. e., Y U T R, called tag assignments (TAS for short), and is a user-specic subtag/supertag-relation, i. e., U T T , called subtag/supertag relation. The personomy Pu of a given user u U is the restriction of F to u, i. e., Pu := (Tu , Ru , Iu , u ) with Iu := {(t, r) T R | (u, t, r) Y }, Tu := 1 (Iu ), Ru := 2 (Iu ), and u := {(t1 , t2 ) T T | (u, t1 , t2 ) }, where i denotes the projection on the ith dimension. Users are typically described by their user ID, and tags may be arbitrary strings. What is considered as a resource depends on the type of system. For instance, in del.icio.us, the resources are URLs, and in ickr, the resources are pictures. From an implementation point of view, resources are internally represented by some ID. In this paper, we do not make use of the subtag/supertag relation for sake of simplicity. I. e., = , and we will simply note a folksonomy as a quadruple F := (U, T, R, Y ). This structure is known in Formal Concept Analysis [14, 4] as a triadic context [7, 13]. An equivalent view on folksonomy data is that of a tripartite (undirected) hypergraph G = (V, E), where V = U T R is the set of nodes, and E = {{u, t, r} | (u, t, r) Y } is the set of hyperedges. 2.3 Del.ico.us A Folksonomy-Based Social Bookmark System In order to evaluate our retrieval technique detailed in the next section, we have analyzed the popular social bookmarking sytem del.icio.us, which is a server-based system with a simple-to-use interface that allows users to organize and share bookmarks on the internet. It is able to store in addition to the URL a description, an extended description, and tags (i. e., arbitrary labels). We chose del.icio.us rather than our own system, BibSonomy, as the latter went online only after the time of writing of this article. For our experiments, we collected data from the del.ico.us system in the following way. Initially we used wget starting from the top page of del.ico.us to obtain nearly 6900 users and 700 tags as a starting set. Out of this dataset we extracted all users and resources (i. e., del.icio.us MD5-hashed urls). From July 27 to 30, 2005, we downloaded in a recursive manner user pages to get new resources, and resource pages to get new

1 "Tags" "Users" "Resources" 0.1

0.01

Percentage

0.001

1e-04

1e-05

1e-06

1e-07 1 10 100 1000 Number of Occurrences 10000 100000 1e+06

Fig. 2. Number of TAS occurrences for tags, users, resources in del.icio.us

users. Furthermore we monitored the del.icio.us start page to gather additional users and resources. This way we collected a list of several thousand usernames which we used for accessing the rst 10000 resources each user had tagged. From the collected data we nally took the user les to extract resources, tags, dates, descriptions, extended descriptions, and the corresponding username. We obtained a core folksonomy with |U | = 75, 242 users, |T | = 533, 191 tags and |R| = 3, 158, 297 resources, related by in total |Y | = 17, 362, 212 TAS.15 After inserting this dataset into a MySQL database, we were able to perform our evaluations, as described in the following sections. As expected, the tagging behavior in del.icio.us shows a power law distribution, see Figure 2. This gure presents the percentage of tags, users, and resources, respectively, which occur in a given number of TAS. For instance, the rightmost + indicates that a fraction of 2.19 106 of all tags (i. e. one tag) occurs 415950 times in this case it is the empty tag. The next + shows that one tag (web) occurs 238891 times, and so on. One observes that while the tags follow a power law distribution very strictly, the plot for users and resources levels off for small numbers of occurrences. Based on this observation, we estimate to have crawled most of the tags, while many users and resources are still missing from the dataset. A probable reason is that many users only try posting a single resource, often without entering any tags (the empty tag is the most frequent one in the dataset), before they decide not to use the system anymore. These users and resources are very unlikely to be connected with others at all (and they only appear for a short period on the del.icio.us start page), so that they are not included in our crawl.

15

4,313 users additionally organised 113,562 of the tags with 6,527 so-called bundles. The bundles will not be discussed in this paper; they can be interpreted as one level of the relation.

3 Ranking in Folksonomies using Adapted PageRank

Current folksonomy tools such as del.icio.us provide only very limited search support in addition to their browsing interface. Searching can be performed over the text of tags and resource descriptions, but no ranking is done apart from ordering the hits in reverse chronological order. Using traditional information retrieval, folksonomy contents can be searched textually. However, as the documents consist of short text snippets only (usually a description, e. g. the web page title, and the tags themselves), ordinary ranking schemes such as TF/IDF are not feasible. As shown in Section 2.2, a folksonomy induces a graph structure which we will exploit for ranking in this section. Our FolkRank algorithm is inspired by the seminal PageRank algorithm [2]. The PageRank weight-spreading approach cannot be applied directly on folksonomies because of the different nature of folksonomies compared to the web graph (undirected triadic hyperedges instead of directed binary edges). In the following we discuss how to overcome this problem. 3.1 Adaptation of PageRank We implement the weight-spreading ranking scheme on folksonomies in two steps. First, we transform the hypergraph between the sets of users, tags, and resources into an undirected, weighted, tripartite graph. On this graph, we apply a version of PageRank that takes into account the edge weights. Converting the Folksonomy into an Undirected Graph. First we convert the folksonomy F = (U, T, R, Y ) into an undirected tripartite graph GF = (V, E) as follows. 1. The set V of nodes of the graph consists of the disjoint union of the sets of tags, users and resources: V = U T R. (The tripartite structure of the graph can be exploited later for an efcient storage of the sparse adjacency matrix and the implementation of the weight-spreading iteration in the FolkRank algorithm.) 2. All co-occurrences of tags and users, users and resources, tags and resources become undirected, weighted edges between the respective nodes: E = {{u, t}, {t, r}, {u, r} | (u, t, r) Y }, with each edge {u, t} being weighted with |{r R : (u, t, r) Y }|, each edge {t, r} with |{u U : (u, t, r) Y }|, and each edge {u, r} with |{t T : (u, t, r) Y }|. Folksonomy-Adapted Pagerank. The original formulation of PageRank [2] reects the idea that a page is important if there many pages linking to it, and if those pages are important themselves. The distribution of weights can thus be described as the xed point of a weight passing scheme on the web graph. This idea was extended in a similar fashion to bipartite subgraphs of the web in HITS [6] and to n-ary directed graphs in [15]). We employ the same underlying principle for our ranking scheme in folksonomies. The basic notion is that a resource which is tagged with important tags by important users becomes important itself. The same holds, symmetrically, for tags and users. Thus we have a graph of vertices which are mutually reinforcing each other by spreading their weights.

Like PageRank, we employ the random surfer model, a notion of importance for web pages that is based on the idea that an idealized random web surfer normally follows hyperlinks, but from time to time randomly jumps to a new webpage without following a link. This results in the following denition of the rank of the vertices of the graph the entries in the xed point w of the weight spreading computation w dAw + (1 d)p, where w is a weight vector with one entry for each web page, A is the row-stochastic16 version of the adjacency matrix of the graph GF dened above, p is the random surfer component, and d [0, 1] is determining the inuence of p. In the original PageRank, p is used to outweigh the loss of weight on web pages without outgoing links. Usually, one will choose p = 1, i. e., the vector composed by 1s. In order to compute personalized PageRanks, however, p can be used to express user preferences by giving a higher weight to the components which represent the users preferred web pages. We employ a similar motivation for our ranking scheme in folksonomies. The basic notion is that a resource which is tagged with important tags by important users becomes important itself. The same holds, symmetrically, for tags and users, thus we have a tripartite graph in which the vertices are mutually reinforcing each other by spreading their weights. Formally, we spread the weight as follows: w w + Aw + p (1)

where A is the row-stochastic version of the adjacency matrix of GF , p is a preference vector, , , [0, 1] are constants with + + = 1. The constant is intended to regulate the speed of convergence, while the proportion between and controls the inuence of the preference vector. We call the iteration according to Equation 1 until convergence is achieved the Adapted PageRank algorithm. Note that, if ||w||1 = ||p||1 holds,17 the sum of the weights in the system will remain constant. The inuence of different settings of the parameters , , and is discussed below. As the graph GF is undirected, part of the weight that went through an edge at moment t will ow back at t + 1. The results are thus rather similar (but not identical) to a ranking that is simply based on edge degrees, as we will see now. The reason for applying the more expensive PageRank approach nonetheless is that its random surfer vector allows for topic-specic ranking, as we will discuss in the next section. 3.2 Results for Adapted PageRank We have evaluated the Adapted PageRank on the del.ico.us dataset described in Section 2.3. As there exists no gold standard ranking on these data, we evaluate our results empirically. First, we studied the speed of convergence. We let p := 1 (the vector having 1 in all components), and varied the parameter settings. In all settings, we discovered that = 0 slows down the convergence rate. For instance, for = 0.35, = 0.65, = 0, 411 iterations were needed, while = 0, = 1, = 0 returned the same result in only 320 iterations. It turns out that using as a damping factor by spreading equal weight
I. e., each row of the matrix is normalized to 1 in the 1-norm. sinks but this holds trivially in our graph GF .
16 17

. . . and if there are no rank

Table 1. Folksonomy Adapted PageRank applied without preferences (called baseline)

ad. PageRank 0,0007389 0,0007379 0,0006796 0,0006171 0,0005044 0,0004885 0,0004844 0,0004820 0,0004797 0,0004668 0,0004556 0,0004543 0,0004535 0,0004520 0,0004355 0,0003829 0,0003727 0,0003600 0,0003462 0,0003419 0,0003407 URL ad. PageRank http://slashdot.org/ 0,0002613 http://pchere.blogspot.com/2005/02/absolutely-delicious-complete-tool.html 0,0002320 http://script.aculo.us/ 0,0001770 http://www.adaptivepath.com/publications/essays/archives/000385.php 0,0001654 http://johnvey.com/features/deliciousdirector/ 0,0001593 http://en.wikipedia.org/wiki/Main Page 0,0001407 http://www.ickr.com/ 0,0001376 http://www.goodfonts.org/ 0,0001349 http://www.43folders.com/ 0,0001160 http://www.csszengarden.com/ 0,0001149 http://wellstyled.com/tools/colorscheme2/index-en.html 0,0001108 http://pro.html.it/esempio/nifty/ 0,0001070 http://www.alistapart.com/ 0,0001059 http://postsecret.blogspot.com/ 0,0001058 http://www.beelerspace.com/index.php?p=890 0,0001035 http://www.techsupportalert.com/best 46 free utilities.htm 0,0001034 http://www.alvit.de/web-dev/ 0,0001020 http://www.technorati.com/ 0,0001015 http://www.lifehacker.com/ 0,0001009 http://www.lucazappa.com/brilliantMaker/buttonImage.php 0,0000992 http://www.engadget.com/ 0,0000984 Tag system:unled web blog design software music programming css reference linux tools news art blogs politics java javascript mac games photography fun ad. PageRank 0,0078404 0,0044031 0,0042003 0,0041828 0,0038904 0,0037273 0,0037100 0,0030766 0,0026019 0,0024779 0,0024147 0,0023611 0,0023358 0,0021035 0,0019371 0,0018757 0,0017610 0,0017252 0,0015801 0,0015469 0,0015296 User shankar notmuch fritz ubi.quito.us weev kof2002 ukquake gearhead angusf johncollins mshook frizzlebiscuit rafaspol xiombarg tidesonar02 cyrusnews bldurling onpause tv anytime cataracte triple entendre kayodeok

to each node in each iteration speeds up the convergence considerably by a factory of approximately 10 (e. g., 39 iterations for = 0, = 0.85, = 0.15). Table 1 shows the result of the adapted PageRank algorithm for the 20 most important tags, users and resources computed with the parameters = 0.35, = 0.65, = 0 (which equals the result for = 0, = 1, = 0). Tags get the highest ranks, followed by the users, and the resources. Therefore, we present their rankings in separate lists. As we can see from the tag table, the most important tag is system:unled which is used to indicate that a user did not assign any tag to a resource. It is followed by web, blog, design etc. This corresponds more or less to the rank of the tags given by the overall tag count in the dataset. The reason is that the graph GF is undirected. We face thus the problem that, in the Adapted PageRank algorithm, weights that ow in one direction of an edge will basically swash back along the same edge in the next

iteration. Therefore the resulting is very similar (although not equal!) to a ranking based on counting edge degrees. The resource ranking shows that Web 2.0 web sites like Slashdot, Wikipedia, Flickr, and a del.icio.us related blog appear in top positions. This is not surprising, as early users of del.ico.us are likely to be interested in Web 2.0 in general. This ranking correlates also strongly with a ranking based on edge counts. The results for the top users are of more interest as different kinds of users appear. As all top users have more than 6000 bookmarks; notmuch has a large amount of tags, while the tag count of fritz is considerably smaller. To see how good the topic-specic ranking by Adapted PageRank works, we combined it with term frequency, a standard information retrieval weighting scheme. To this end, we downloaded all 3 million web pages referred to by a URL in our dataset. From these, we considered all plain text and html web pages, which left 2.834.801 documents. We converted all web pages into ASCII and computed an inverted index. To search for a term as in a search engine, we retrieved all pages containing the search term and ranked them by tf(t) w[v] where tf(t) is the term frequency of search term t in page v, and w[v] is the Adapted PageRank weight of v. Although this is a rather straightforward combination of two successful retrieval techniques, our experiments with different topic-specic queries indicate that this adaptation of PageRank does not work very well. For instance, for the search term football, the del.icio.us homepage showed up as the rst result. Indeed, most of the highly ranked pages have nothing to do with football. Other search terms provided similar results. Apparently, the overall structure of the undirected graph overrules the inuence of the preference vector. In the next section, we discuss how to overcome this problem.

4 FolkRank Topic-Specic Ranking in Folksonomies

In order to reasonably focus the ranking around the topics dened in the preference vector, we have developed a differential approach, which compares the resulting rankings with and without preference vector. This resulted in our new FolkRank algorithm. 4.1 The FolkRank Algorithm The FolkRank algorithm computes a topic-specic ranking in a folksonomy as follows: 1. The preference vector p is used to determine the topic. It may have any distribution of weights, as long as ||w||1 = ||p||1 holds. Typically a single entry or a small set of entries is set to a high value, and the remaining weight is equally distributed over the other entries. Since the structure of folksonomies is symmetric, we can dene a topic by assigning a high value to either one or more tags and/or one or more users and/or one or more resources. 2. Let w0 be the xed point from Equation (1) with = 1. 3. Let w1 be the xed point from Equation (1) with < 1. 4. w := w1 w0 is the nal weight vector.

Thus, we compute the winners and losers of the mutual reinforcement of resources when a user preference is given, compared to the baseline without a preference vector. We call the resulting weight w[x] of an element x of the folksonomy the FolkRank of x. Whereas the Adapted PageRank provides one global ranking, independent of any preferences, FolkRank provides one topic-specic ranking for each given preference vector. Note that a topic can be dened in the preference vector not only by assigning higher weights to specic tags, but also to specic resources and users. These three dimensions can even be combined in a mixed vector. Similarly, the ranking is not restricted to resources, it may as well be applied to tags and to users. We will show below that indeed the rankings on all three dimensions provide interesting insights. 4.2 Comparing FolkRank with Adapted PageRank To analyse the proposed FolkRank algorithm, we generated rankings for several topics, and compared them with the ones obtained from Adapted PageRank. We will here discuss two sets of search results, one for the tag boomerang, and one for the URL http.//www.semanticweb.org. Our other experiments all provided similar results. The leftmost part of Table 2 contains the ranked list of tags according to their weights from the Adapted PageRank by using the parameters = 0.2, = 0.5, = 0.3, and 5 as a weight for the tag boomerang in the preference vector p, while the other elements were given a weight of 0. As expected, the tag boomerang holds the rst position while tags like shop or wood which are related are also under the Top 20. The tags software, java, programming or web, however, are on positions 4 to 7, but have nothing to do with boomerang. The only reason for their showing up is that they are frequently used in del.icio.us (cf. Table 1). The second column from the left in Table 2 contains the results of our FolkRank algorithm, again for the tag boomerang. Intuitively, this ranking is better, as the globally frequent words disappear and related words like wood and construction are ranked higher. A closer look reveals that this ranking still contains some unexpected tags; kassel or rdf are for instance not obviously related to boomerang. An analysis of the user ranking (not displayed) explains this fact. The top-ranked user is schm4704, and he has indeed many bookmarks about boomerangs. A FolkRank run with preference weight 5 for user schm4704 shows his different interests, see the rightmost column in Table 2. His main interest apparently is in boomerangs, but other topics show up as well. In particular, he has a strong relationship to the tags kassel and rdf. When a community in del.ico.us is small (such as the boomerang community), already a single user can thus provide a strong bridge to other communities, a phenomenon that is equally observed in small social communities. A comparison of the FolkRank ranking for user schm4704 with the Adapted PageRank result for him (2nd ranking from left) conrms the initial nding from above, that the Adapted PageRank ranking contains many globally frequent tags, while the FolkRank ranking provides more personal tags. While the differential nature of the FolkRank algorithm usually pushes down the globally frequent tags such as web, though, this happens in a differentiated manner: FolkRank will keep them in the top positions, if they are indeed relevant to the user under consideration. This can be seen for example for the tags web and java. While the tag web appears in schm4704s tag list but not very of-

Table 2. Ranking results for the tag boomerang (two left at top: Adapted PageRank and FolkRank for tags, middle: FolkRank for URLs) and for the user schm4704 (two right at top: Adapted PageRank and FolkRank for tags, bottom: FolkRank for URLs)
Tag ad. PRank Tag FolkRank Tag ad. PRank Tag FolkRank boomerang 0,4036883 boomerang 0,4036867 boomerang 0,0093549 boomerang 0,0093533 shop 0,0069058 shop 0,0066477 lang:ade 0,0068111 lang:de 0,0068028 lang:de 0,0050943 lang:de 0,0050860 shop 0,0052600 shop 0,0050019 software 0,0016797 wood 0,0012236 java 0,0052050 java 0,0033293 java 0,0016389 kassel 0,0011964 web 0,0049360 kassel 0,0032223 programming 0,0016296 construction 0,0010828 programming 0,0037894 network 0,0028990 web 0,0016043 plans 0,0010085 software 0,0035000 rdf 0,0028758 reference 0,0014713 injuries 0,0008078 network 0,0032882 wood 0,0028447 system:unled 0,0014199 pitching 0,0007982 kassel 0,0032228 delicious 0,0026345 wood 0,0012378 rdf 0,0006619 reference 0,0030699 semantic 0,0024736 kassel 0,0011969 semantic 0,0006533 rdf 0,0030645 database 0,0023571 linux 0,0011442 material 0,0006279 delicious 0,0030492 guitar 0,0018619 construction 0,0011023 triy 0,0005691 system:unled 0,0029393 computing 0,0018404 plans 0,0010226 network 0,0005568 linux 0,0029393 cinema 0,0017537 network 0,0009460 webring 0,0005552 wood 0,0028589 lessons 0,0017273 rdf 0,0008506 sna 0,0005073 database 0,0026931 social 0,0016950 css 0,0008266 socialnetworkanalysis 0,0004822 semantic 0,0025460 documentation 0,0016182 design 0,0008248 cinema 0,0004726 css 0,0024577 scientic 0,0014686 delicious 0,0008097 erie 0,0004525 social 0,0021969 lesystem 0,0014212 injuries 0,0008087 riparian 0,0004467 webdesign 0,0020650 userspace 0,0013490 pitching 0,0007999 erosion 0,0004425 computing 0,0020143 library 0,0012398 Url FolkRank http://www.ight-toys.com/boomerangs.htm 0,0047322 http://www.ight-toys.com/ 0,0047322 http://www.bumerangclub.de/ 0,0045785 http://www.bumerangbel.de/ 0,0045781 http://www.kutek.net/triy mods.php 0,0032643 http://www.rediboom.de/ 0,0032126 http://www.bws-buhmann.de/ 0,0032126 http://www.akspiele.de/ 0,0031813 http://www.medco-athletics.com/education/elbow shoulder injuries/ 0,0031606 http://www.sportsprolo.com/sports%20prolotherapy%20newsletter%20pitching%20injuries.htm 0,0031606 http://www.boomerangpassion.com/english.php 0,0031005 http://www.kuhara.de/bumerangschule/ 0,0030935 http://www.bumerangs.de/ 0,0030935 http://s.webring.com/hub?ring=boomerang 0,0030895 http://www.kutek.net/boomplans/plans.php 0,0030873 http://www.geocities.com/cmorris32839/jonas article/ 0,0030871 http://www.theboomerangman.com/ 0,0030868 http://www.boomerangs.com/index.html 0,0030867 http://www.lmifox.com/us/boom/index-uk.htm 0,0030867 http://www.sports-boomerangs.com/ 0,0030867 http://www.rangsboomerangs.com/ 0,0030867 Url FolkRank http://jena.sourceforge.net/ 0,0019369 http://www.openrdf.org/doc/users/ch06.html 0,0017312 http://dsd.lbl.gov/ hoschek/colt/api/overview-summary.html 0,0016777 http://librdf.org/ 0,0014402 http://www.hpl.hp.com/semweb/jena2.htm 0,0014326 http://jakarta.apache.org/commons/collections/ 0,0014203 http://www.aktors.org/technologies/ontocopi/ 0,0012839 http://eventseer.idi.ntnu.no/ 0,0012734 http://tangra.si.umich.edu/ radev/ 0,0012685 http://www.cs.umass.edu/ mccallum/ 0,0012091 http://www.w3.org/TR/rdf-sparql-query/ 0,0011945 http://ourworld.compuserve.com/homepages/graeme birchall/HTM COOK.HTM 0,0011930 http://www.emory.edu/EDUCATION/mfp/Kuhn.html 0,0011880 http://www.hpl.hp.com/semweb/rdql.htm 0,0011860 http://jena.sourceforge.net/javadoc/index.html 0,0011860 http://www.geocities.com/mailsoftware42/db/ 0,0011838 http://www.quirksmode.org/ 0,0011327 http://www.kde.cs.uni-kassel.de/lehre/ss2005/googlespam 0,0011110 http://www.powerpage.org/cgi-bin/WebObjects/powerpage.woa/wa/story?newsID=14732 0,0010402 http://www.vaughns-1-pagers.com/internet/google-ranking-factors.htm 0,0010329 http://www.cl.cam.ac.uk/Research/SRG/netos/xen/ 0,0010326

ten, java is a very important tag for that user. This is reected in the FolkRank ranking: java remains in the Top 5, while web is pushed down in the ranking. The ranking of the resources for the tag boomerang given in the middle of Table 2 also provides interesting insights. As shown in the table, many boomerang related web pages show up (their topical relatedness was conrmed by a boomerang acionado). Comparing the Top 20 web pages of boomerang with the Top 20 pages given by the schm4704 ranking, there is no boomerang web page in the latter. This can be explained by analysing the tag distribution of this user. While boomerang is the most frequent tag for this user, in del.icio.us, boomerang appears rather infrequently. The rst boomerang web page in the schm4704 ranking is the 21st URL (i. e., just outside the listed TOP 20). Thus, while the tag boomerang itself dominates the tags of this user, in the whole, the semantic web related tags and resources prevail. This demonstrates that while the user schm4704 and the tag boomerang are strongly correlated, we can still get an overview of the respective related items which shows several topics of interest for the user. Let us consider a second example. Table 3 gives the results for the web page http://www.semanticweb.org/. The two tables on the left show the tags and users for the adapted PageRank, resp., and the two ones on the right the FolkRank results. Again, we see that the differential ranking of FolkRank makes the right decisions: in the Adaptive PageRank, globally frequent tags such as web, css, xml, programming get high ranks. Of these, only two turn up to be of genuine interest to the members of the Semantic Web community: web and xml remain at high positions, while css and programming disappear altogether from the list of the 20 highest ranked tags. Also, several variations of tags which are used to label Semantic Web related pages appear (or get ranked higher): semantic web (two tags, space-separated), semantic web, semweb, sem-web. These co-occurrences of similar tags could be exploited further to consolidate the emergent semantics of a eld of interest. While the discovery in this case may also be done in a simple syntactic analysis, the graph based approach allows also for detecting inter-community and inter-language relations. The user IDs can not be checked for topical relatedness immediately, since they are not related to the users full names although a former winner of the Semantic Web Challenge and the best paper award at a Semantic Web Conference seems to be among them. The web pages that appear in the top list, on the other hand, include many well-known resources from the Semantic Web area. An interesting resource on the list is PiggyBank, which has been presented in November 2005 at the ISWC conference. Considering that the dataset was crawled in July 2005, when PiggyBank was not that well known, the prominent position of PiggyBank in del.icio.us at such an early time is an interesting result. This indicates the sensibility of social bookmarking systems for upcoming topics. These two examples as well as the other experiments we performed show that FolkRank provides good results when querying the folksonomy for topically related elements. Overall, our experiments indicate that topically related items can be retrieved with FolkRank for any given set of highlighted tags, users and/or resources. Our results also show that the current size of folksonomies is still prone to being skewed by a relatively small number of perturbations a single user, at the moment, can inuence the emergent understanding of a certain topic in the case that a sufcient number of different points of view for such a topic has not been collected yet. With the

Table 3. Ranking for the resource http://www.semanticweb.org (Left two tables: Adapted PageRank for tags and users; right two tables: FolkRank for tags and users. Bottom: FolkRank for resources).
Tag semanticweb web semantic system:unled semantic web rdf semweb resources community xml research programming css portal .imported imported-bo... en science .idate2005-04-11 newfurl internet ad. PRank User ad. PageRank Tag FolkRank User 0,0208605 up4 0,0091995 semanticweb 0,0207820 up4 0,0162033 awenger 0,0086261 semantic 0,0121305 awenger 0,0122028 j.deville 0,0074021 web 0,0118002 j.deville 0,0088625 chaizzilla 0,0062570 semantic web 0,0071933 chaizzilla 0,0072150 elektron 0,0059457 rdf 0,0044461 elektron 0,0046348 captsolo 0,0055671 semweb 0,0039308 captsolo 0,0039897 stevag 0,0049923 resources 0,0034209 dissipative 0,0037884 dissipative 0,0049647 community 0,0033208 stevag 0,0037256 krudd 0,0047574 portal 0,0022745 krudd 0,0031494 williamteo 0,0037204 xml 0,0022074 williamteo 0,0026720 stevecassidy 0,0035887 research 0,0020378 stevecassidy 0,0025717 pmika 0,0035359 imported-bo... 0,0018920 pmika 0,0025290 millette 0,0033028 en 0,0018536 millette 0,0024118 myren 0,0028117 .idate2005-04-11 0,0017555 myren 0,0020495 morningboat 0,0025913 newfurl 0,0017153 morningboat 0,0019610 philip.fennell 0,0025338 tosort 0,0014486 philip.fennell 0,0018900 mote 0,0025212 cs 0,0014002 webb. 0,0018166 dnaboy76 0,0024813 academe 0,0013822 dnaboy76 0,0017779 webb. 0,0024709 rd 0,0013456 mote 0,0017578 nymetbarton 0,0023790 sem-web 0,0013316 alphajuliet 0,0016122 alphajuliet 0,0023781 w3c 0,0012994 nymetbarton URL FolkRank http://www.semanticweb.org/ 0,3761957 http://ink.semanticweb.org/ 0,0005566 http://simile.mit.edu/piggy-bank/ 0,0003828 http://www.w3.org/2001/sw/ 0,0003216 http://infomesh.net/2001/swintro/ 0,0002162 http://del.icio.us/register 0,0001745 http://mspace.ecs.soton.ac.uk/ 0,0001712 http://www.adaptivepath.com/publications/essays/archives/000385.php 0,0001637 http://www.ontoweb.org/ 0,0001617 http://www.aaai.org/AITopics/html/ontol.html 0,0001613 http://simile.mit.edu/ 0,0001395 http://itip.evcc.jp/itipwiki/ 0,0001256 http://www.google.be/ 0,0001224 http://www.letterjames.de/index.html 0,0001224 http://www.daml.org/ 0,0001216 http://shirky.com/writings/ontology overrated.html 0,0001195 http://jena.sourceforge.net/ 0,0001167 http://www.alistapart.com/ 0,0001102 http://www.federalconcierge.com/WritingBusinessCases.html 0,0001060 http://pchere.blogspot.com/2005/02/absolutely-delicious-complete0,0001059 tool.html http://www.shirky.com/writings/semantic syllogism.html 0,0001052 FolkRank 0,0091828 0,0084958 0,0073525 0,0062227 0,0059403 0,0055369 0,0049619 0,0049590 0,0047005 0,0037181 0,0035840 0,0035358 0,0032103 0,0027965 0,0025875 0,0025145 0,0024671 0,0024659 0,0024214 0,0023668 0,0023666

growth of folksonomy-based data collections on the web, the inuence of single users will fade in favor of a common understanding provided by huge numbers of users. As detailed above, our ranking is based on tags only, without regarding any inherent features of the resources at hand. This allows to apply FolkRank to search for pictures (e. g., in ickr) and other multimedia content, as well as for all other items that are difcult to search in a content-based fashion. The same holds for intranet applications, where in spite of centralized knowledge management efforts, documents often remain unused because they are not hyperlinked and difcult to nd. Full text retrieval may be used to nd documents, but traditional IR methods for ranking without hyperlink information have difculties nding the most relevant documents from large corpora.

4.3 Generating Recommendations The original PageRank paper [2] already pointed out the possibility of using the random surfer vector p as a personalization mechanism for PageRank computations. The results of Section 4 show that, given a user, one can nd set of tags and resources of interest to him. Likewise, FolkRank yields a set of related users and resources for a given tag. Following these observations, FolkRank can be used to generate recommendations within a folksonomy system. These recommendations can be presented to the user at different points in the usage of a folksonomy system: Documents that are of potential interest to a user can be suggested to him. This kind of recommendation pushes potentially useful content to the user and increases the chance that a user nds useful resources that he did not even know existed by serendipitous browsing. When using a certain tag, other related tags can be suggested. This can be used, for instance, to speed up the consolidation of different terminologies and thus facilitate the emergence of a common vocabulary. While folksonomy tools already use simple techniques for tag recommendations, FolkRank additionally considers the tagging behavior of other users. Other users that work on related topics can be made explicit, improving thus the knowledge transfer within organizations and fostering the formation of communities.

5 Conclusion and Outlook

In this paper, we have argued that enhanced search facilities are vital for emergent semantics within folksonomy-based systems. We presented a formal model for folksonomies, the FolkRank ranking algorithm that takes into account the structure of folksonomies, and evaluation results on a large-scale dataset. The FolkRank ranking scheme has been used in this paper to generate personalized rankings of the items in a folksonomy, and to recommend users, tags and resources. We have seen that the top folksonomy elements which are retrieved by FolkRank tend to fall into a coherent topic area, e.g. Semantic Web. This leads naturally to the idea of extracting communities of interest from the folksonomy, which are represented by their top tags and the most inuential persons and resources. If these communities are made explicit, interested users can nd them and participate, and community members can more easily get to know each other and learn of others resources. Another future research issue is to combine different search and ranking paradigms. In this paper, we went a rst step by focusing on the new structure of folksonomies. In the future, we will incorporate additionally the full text that is contained in the web pages addressed by the URLs, the link structure of these web pages, and the usage behavior as stored in the log le of the tagging system. The next version will also exploit the tag hierarchy. Currently, spam is not a serious problem for social bookmarking systems. With the increasing attention they currently receive, however, we anticipate that spam posts will show up sooner or later. As for mail spam and link farms in the web, solutions will be needed to lter out spam. We expect that a blend of graph structure analysis together with content analysis will give the best results.

When folksonomy-based systems grow larger, user support has to go beyond enhanced retrieval facilities. Therefore, the internal structure has to become better organized. An obvious approach for this are semantic web technologies. The key question remains though how to exploit its benets without bothering untrained users with its rigidity. We believe that this will become a fruitful research area for the Semantic Web community for the next years. Acknowledgement. Part of this research was funded by the EU in the Nepomuk project (FP6-027705).

References
1. Harith Alani, Srinandan Dasmahapatra, Kieron OHara, and Nigel Shadbolt. Identifying Communities of Practice through Ontology Network Analysis. IEEE Intelligent Systems, 18(2):18 25, March/April 2003. 2. Sergey Brin and Lawrence Page. The Anatomy of a Large-Scale Hypertextual Web Search Engine. Computer Networks and ISDN Systems, 30(1-7):107117, April 1998. 3. Connotea Mailing List. https://lists.sourceforge.net/lists/listinfo/connotea-discuss. 4. B. Ganter and R. Wille. Formal Concept Analysis : Mathematical foundations. Springer, 1999. 5. Tony Hammond, Timo Hannay, Ben Lund, and Joanna Scott. Social Bookmarking Tools (I): A General Review. D-Lib Magazine, 11(4), April 2005. 6. Jon M. Kleinberg. Authoritative sources in a hyperlinked environment. Journal of the ACM, 46(5):604632, 1999. 7. F. Lehmann and R. Wille. A triadic approach to formal concept analysis. In G. Ellis, R. Levinson, W. Rich, and J. F. Sowa, editors, Conceptual Structures: Applications, Implementation and Theory, volume 954 of Lecture Notes in Computer Science. Springer, 1995. 8. Ben Lund, Tony Hammond, Martin Flack, and Timo Hannay. Social Bookmarking Tools (II): A Case Study - Connotea. D-Lib Magazine, 11(4), April 2005. 9. Adam Mathes. Folksonomies Cooperative Classication and Communication Through Shared Metadata, December 2004. http://www.adammathes.com/academic/computermediated-communication/folksonomies.html. 10. Peter Mika. Ontologies Are Us: A Unied Model of Social Networks and Semantics. In Yolanda Gil, Enrico Motta, V. Richard Benjamins, and Mark A. Musen, editors, ISWC 2005, volume 3729 of LNCS, pages 522536, Berlin Heidelberg, November 2005. Springer-Verlag. 11. S. Staab, S. Santini, F. Nack, L. Steels, and A. Maedche. Emergent semantics. Intelligent Systems, IEEE [see also IEEE Expert], 17(1):7886, 2002. 12. L. Steels. The origins of ontologies and communication conventions in multi-agent systems. Autonomous Agents and Multi-Agent Systems, 1(2):169194, October 1998. 13. Gerd Stumme. A nite state model for on-line analytical processing in triadic contexts. In Bernhard Ganter and Robert Godin, editors, ICFCA, volume 3403 of Lecture Notes in Computer Science, pages 315328. Springer, 2005. 14. R. Wille. Restructuring lattices theory : An approach based on hierarchies of concepts. In I. Rival, editor, Ordered Sets, pages 445470. Reidel, Dordrecht-Boston, 1982. 15. W. Xi, B. Zhang, Y. Lu, Z. Chen, S. Yan, H. Zeng, W. Ma, and E. Fox. Link fusion: A unied link analysis framework for multi-type interrelated data objects. In Proc. 13th Internation World Wide Web Conference, New York, 2004.