A Tag Recommender System Exploiting User and Community Behavior Cataldo musto Fedelucio narducci Marco de gemmis Dept of Computer Science Dept. of Computer Science Dept. of Computer Science University of Bari 'Aldo Moro University of Bari Aldo Moro University of Bari 'Aldo Moro cataldomusto@di uniba. it narducci@di uniba. it degemmis @di uniba. it Pasquale Lops Giovanni Semeraro ept. of Computer Dept of Computer Science University of Bari Aldo Moro University of Bari 'Aldo Moro Italy lops @di uniba. it semeraro@diuniba. it ABSTRACT Information filtering Nowadays Web sites tend to be more and more social: users can upload any kind of information on collaborative plat- General Terms forms and can express their opinions about the content they njoyed through textual feedbacks or reviews. These plat Algorithms, Experimentation forms allow users to annotate resources they like through reely chosen keywords(called tags). The main advantage Ke of these tools is that they perfectly fit user needs, since the Recommender Systems, Web 2.0, Collaborative Tagging Sys- Ise of tags allows organizing the information in a way that closely follows the user mental model, making retrieval of tems, Folksonomies formation easier. However, the heterogeneity characteriz- ing the communities causes some problems in the activity of 1. INTRODUCTION social tagging: someone annotates resources with very sp We are assisting to a transformation of the Web towards cific tags, other people with generic ones, and so on. These a more user-centric vision called Web 2.0. By using Web 2.0 drawbacks reduce the exploitation of collaborative tagging plications users are able to publish auto-produced con- systems for retrieval and filtering tasks. Therefore, systems tents such as photos, videos, political opinions, reviews, hence hat assist the user in the task of tagging are required. The goal of these systems, called tag recommenders, is to suggest of knowledge. Recently the research community has lloers ey are identified as Web prosumers: producers +consum a set of relevant keywords for the resources to be annotated. oughly analyzed the dynamics of tagging, which is the act This paper presents a tag recommender system called STar of annotating resources with free labels, called tags. any (Social Tag Recommender system). Our system is based on argue that, thanks to the expressive power of folksonomies two assumptions: 1) the more two or more resources are si 17, collaborative tagging systems are very helpful to users ar, the more they share common tags 2)a tag recommender in organizing, browsing and searching resources. This hap- should be able to exploit tags the user already used in order pens because, in contrast to systems where information about to extract useful keywords to label new resources. We also resources is only provided by a small set of experts, the present an experimental evaluation carried out using a large model of collaborative tagging systems takes into account dataset gathered from Bibsonomy. the way individuals conceive the information contained in a resource[18, so they perfectly fit user needs and user mental Categories and Subject Descriptors model. Nowadays almost all Web 2.0 platforms embed tag H 3.1 [Information Storage and Retrieval]: Content Bibsonomy and so on. These systems provide heteroge- Analysis and Indexing: Indexing methods; H.3.3 Information neous contents(photos, videos, musical habits, etc. ), but Storage and Retrieval]: Information Search and Retrieval hey all share a common core: they let users to post new re- sources and to annotate them with tags. Besides the simple act of annotation, the tagging of resources has also a key so- cial aspect; the connection between users, resources and tags generates a tripartite graph that can be easily exploited t Permission to make digital or hard copies of all or part of this work for se is granted without http://www.fickr.com ot made or distributed for profit or commercial advantage and that copie bear this notice and the full citation on the first page. To copy otherwise, http://www.youtube.com publish, to post on servers or to redistribute to lists, requires prior specific http://delicious.com http://www.last.fml Copyright 200X ACM X-XXXXX-XX-X/XX/XX.$10.00. tp://www.bibsonomy.org/
A Tag Recommender System Exploiting User and Community Behavior Cataldo Musto Dept. of Computer Science University of Bari ‘Aldo Moro’ Italy cataldomusto@di.uniba.it Fedelucio Narducci Dept. of Computer Science University of Bari ‘Aldo Moro’ Italy narducci@di.uniba.it Marco De Gemmis Dept. of Computer Science University of Bari ‘Aldo Moro’ Italy degemmis@di.uniba.it Pasquale Lops Dept. of Computer Science University of Bari ‘Aldo Moro’ Italy lops@di.uniba.it Giovanni Semeraro Dept. of Computer Science University of Bari ‘Aldo Moro’ Italy semeraro@di.uniba.it ABSTRACT Nowadays Web sites tend to be more and more social: users can upload any kind of information on collaborative platforms and can express their opinions about the content they enjoyed through textual feedbacks or reviews. These platforms allow users to annotate resources they like through freely chosen keywords (called tags). The main advantage of these tools is that they perfectly fit user needs, since the use of tags allows organizing the information in a way that closely follows the user mental model, making retrieval of information easier. However, the heterogeneity characterizing the communities causes some problems in the activity of social tagging: someone annotates resources with very specific tags, other people with generic ones, and so on. These drawbacks reduce the exploitation of collaborative tagging systems for retrieval and filtering tasks. Therefore, systems that assist the user in the task of tagging are required. The goal of these systems, called tag recommenders, is to suggest a set of relevant keywords for the resources to be annotated. This paper presents a tag recommender system called STaR (Social Tag Recommender system). Our system is based on two assumptions: 1) the more two or more resources are similar, the more they share common tags 2) a tag recommender should be able to exploit tags the user already used in order to extract useful keywords to label new resources. We also present an experimental evaluation carried out using a large dataset gathered from Bibsonomy. Categories and Subject Descriptors H.3.1 [Information Storage and Retrieval]: Content Analysis and Indexing: Indexing methods; H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval: 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, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Copyright 200X ACM X-XXXXX-XX-X/XX/XX ...$10.00. Information filtering General Terms Algorithms, Experimentation Keywords Recommender Systems, Web 2.0, Collaborative Tagging Systems, Folksonomies 1. INTRODUCTION We are assisting to a transformation of the Web towards a more user-centric vision called Web 2.0. By using Web 2.0 applications users are able to publish auto-produced contents such as photos, videos, political opinions, reviews, hence they are identified as Web prosumers: producers + consumers of knowledge. Recently the research community has thoroughly analyzed the dynamics of tagging, which is the act of annotating resources with free labels, called tags. Many argue that, thanks to the expressive power of folksonomies [17], collaborative tagging systems are very helpful to users in organizing, browsing and searching resources. This happens because, in contrast to systems where information about resources is only provided by a small set of experts, the model of collaborative tagging systems takes into account the way individuals conceive the information contained in a resource [18], so they perfectly fit user needs and user mental model. Nowadays almost all Web 2.0 platforms embed tagging: we can cite Flickr1 , YouTube2 , Del.icio.us3 , Last.fm4 , Bibsonomy5 and so on. These systems provide heterogeneous contents (photos, videos, musical habits, etc.), but they all share a common core: they let users to post new resources and to annotate them with tags. Besides the simple act of annotation, the tagging of resources has also a key social aspect; the connection between users, resources and tags generates a tripartite graph that can be easily exploited to 1 http://www.flickr.com 2http://www.youtube.com 3http://delicious.com/ 4http://www.last.fm/ 5 http://www.bibsonomy.org/
analyze the dynamics of collaborative tagging The paper is organized as follows. Section 2 analyzes re- xample, users that label the same resource lated work. Section 3 explains the architecture of the system came tags might have similar tastes and items and how the recommendation approach is implemented. The e same tags might share common characteristics. experimental evaluation carried out is described in Section Undoubtedly the power of tagging lies in the ability for 4, while conclusions and future work are drawn in the last people to freely determine the appropriate tags for a section source [10. Since folksonomies do not rely on a predefined lexicon or hierarchy they have the main advantage to be 2. RELATED WORK fully free, but at the same time they generate a very noisy Usually the works in the tag recommendation area are tag space, really hardly to exploit for retrieval or recom- broadly divided into three classes: content-based, collabora- mendation tasks without performing any form of processing Golder et. al. 4 identified three major problems of co laborative tagging systems: polysemy, synonymy, and level In the content-based approach, exploiting some Informa- wariation. Polysemy refers to situations where tags can have tion Retrieval-related techniques, a system is able to ex- tract relevant unigrams or bigrams from the text. Brooks et multiple meanings: for example a resource tagged with the al [2], for example, develop a tag recommender system that term bat could indicate a news taken from an online sport newspaper or a Wikipedia article about nature. We refer to exploits TF/IDF scoring [13 in order to automatically sug- gests tags for a blog post. In 5 is presented a novel method ynonymy when multiple tags share a single meaning: for for key term extraction from text documents. Firstly,ev- as'Al', 'artificiaL intelligenceand so on to identify a scien- ery document is modeled as a graph which nodes are terms tific publication about Artificial Intelligence)but also lexical and edges represent semantic relationship between them relations(like resources tagged with arts' 'cultural These graphs are then partitioned using communities de- heritage'). At last, the phenomenon of tagging at different tection techniques and weighted exploiting information ex- levels of abstraction is defined as level-variation. This hap- tracted from Wikipedia. The tags composing the most rel- pens when people annotate the same web page, containing evant communities(a set of terms related with the topic of for example a recipe for roast turkey with the tag roast- the resource) are then suggested to the user turkeybut also with a simple recipe AutoTag [11 is one of the most important systems imple- Since these problems are of hindrance to completely ex menting the collaborative approach for tag recommendation the expressive power of folksonomies, in the last year It presents some analogies with collaborative filtering meth- tools have been developed to assist the user in the ds: as in the collaborative recommender systems the rec task of tagging and to aid at the same time the tag con- ommendations are generated based on the ratings provided vergence 3: we refer to them as tag recommenders. These by similar users(called neighbors), in AutoTag the system suggests tags based on the other tags associated with ilar posts. Firstly, the tool exploits some IR techniques in 1. a user posts a resource order to find similar posts and extracts the tags they are annotated with. All the tags are then merged, building a 2. depending on the approach, the tag recommender ana folksonomy that is filtered and re-ranked. The top-ranked lyzes some information related to the resource(usually tags are finally suggested to the user, who selects the most etadata or a subset of the relations in the aforemen- appropriate ones to attach to the post tioned tripartite graph) Tag Assist [15 extends the Auto Tags approach introduc- 3. the tag recommender processes this information and ing some preprocessing step(specifically, a lossless compres- produces a list of recommended tags sion over existing data) in order to improve the quality of the recommendations. The core of this approach is repre 4. the user freely chooses the most appropriate tags to sented by aa Tag Suggestion Engine(TSE)which leverages previously tagged posts providing appropriate suggestions Clearly, the more these recommended tags match the user Marinho 9 investigates the user-based collaborative ap- needs and her mental model, the more she will use them to proach for tag recommendation. The main outcome of this notate the resource. In this way we can rapidly speed up work shows that users with a similar tag vocabulary tend to the tag convergence aiding at the same time in filtering the. tag alike, since the method seems to produce good results noise of the complete tag space. when applied on the user-tag matrix. This paper presents the tag recommender Star. When de- The problem of tag recommendation through graph-based veloping the model, we tried to point out two concepts: approaches has been firstly addressed by Jaschke et al. in 7 eir Folk Rank algorithm resources with similar content should source which is tagged by important tags from important sers becomes important itself. So, they build a graph a tag recommender needs to take into account the pre- whose nodes mutually reinforces themselves by spreading vious tagging activity of users, increasing the weight their weights. They compared some recommendation tech- of the tags already used to annotate similar resources. niques including collaborative filtering, PageRank and FolkRank, showing that the Folk Rank algorithm outperforms other ap In this work, we identify two main aspects in the tag rec- proaches. Furthermore, Schmitz et al. [14 proposed associ- ommendation task: first, each user has a typical manner ation rule mining as a technique that might be useful in to label resources; second, similar resources usually share tag recommendation process common tags In literature we can find also other methods(called hy-
analyze the dynamics of collaborative tagging systems. For example, users that label the same resource by using the same tags might have similar tastes and items labeled with the same tags might share common characteristics. Undoubtedly the power of tagging lies in the ability for people to freely determine the appropriate tags for a resource [10]. Since folksonomies do not rely on a predefined lexicon or hierarchy they have the main advantage to be fully free, but at the same time they generate a very noisy tag space, really hardly to exploit for retrieval or recommendation tasks without performing any form of processing. Golder et. al. [4] identified three major problems of collaborative tagging systems: polysemy, synonymy, and level variation. Polysemy refers to situations where tags can have multiple meanings: for example a resource tagged with the term bat could indicate a news taken from an online sport newspaper or a Wikipedia article about nature. We refer to synonymy when multiple tags share a single meaning: for example we can have simple morphological variations (such as ’AI’,’artificial intelligence’ and so on to identify a scientific publication about Artificial Intelligence) but also lexical relations (like resources tagged with ‘arts’ versus ‘cultural heritage’). At last, the phenomenon of tagging at different levels of abstraction is defined as level-variation. This happens when people annotate the same web page, containing for example a recipe for roast turkey with the tag ‘roastturkey’ but also with a simple ‘recipe’. Since these problems are of hindrance to completely exploit the expressive power of folksonomies, in the last years many tools have been developed to assist the user in the task of tagging and to aid at the same time the tag convergence [3]: we refer to them as tag recommenders. These systems work in a very simple way: 1. a user posts a resource; 2. depending on the approach, the tag recommender analyzes some information related to the resource (usually metadata or a subset of the relations in the aforementioned tripartite graph); 3. the tag recommender processes this information and produces a list of recommended tags; 4. the user freely chooses the most appropriate tags to annotate the resource. Clearly, the more these recommended tags match the user needs and her mental model, the more she will use them to annotate the resource. In this way we can rapidly speed up the tag convergence aiding at the same time in filtering the noise of the complete tag space. This paper presents the tag recommender STaR. When developing the model, we tried to point out two concepts: • resources with similar content should be annotated with similar tags; • a tag recommender needs to take into account the previous tagging activity of users, increasing the weight of the tags already used to annotate similar resources. In this work, we identify two main aspects in the tag recommendation task: first, each user has a typical manner to label resources; second, similar resources usually share common tags. The paper is organized as follows. Section 2 analyzes related work. Section 3 explains the architecture of the system and how the recommendation approach is implemented. The experimental evaluation carried out is described in Section 4, while conclusions and future work are drawn in the last section. 2. RELATED WORK Usually the works in the tag recommendation area are broadly divided into three classes: content-based, collaborative and graph-based approaches. In the content-based approach, exploiting some Information Retrieval-related techniques, a system is able to extract relevant unigrams or bigrams from the text. Brooks et. al [2], for example, develop a tag recommender system that exploits TF/IDF scoring [13] in order to automatically suggests tags for a blog post. In [5] is presented a novel method for key term extraction from text documents. Firstly, every document is modeled as a graph which nodes are terms and edges represent semantic relationship between them. These graphs are then partitioned using communities detection techniques and weighted exploiting information extracted from Wikipedia. The tags composing the most relevant communities (a set of terms related with the topic of the resource) are then suggested to the user. AutoTag [11] is one of the most important systems implementing the collaborative approach for tag recommendation. It presents some analogies with collaborative filtering methods: as in the collaborative recommender systems the recommendations are generated based on the ratings provided by similar users (called neighbors), in AutoTag the system suggests tags based on the other tags associated with similar posts. Firstly, the tool exploits some IR techniques in order to find similar posts and extracts the tags they are annotated with. All the tags are then merged, building a folksonomy that is filtered and re-ranked. The top-ranked tags are finally suggested to the user, who selects the most appropriate ones to attach to the post. TagAssist [15] extends the AutoTags’ approach introducing some preprocessing step (specifically, a lossless compression over existing data) in order to improve the quality of the recommendations. The core of this approach is represented by a a Tag Suggestion Engine (TSE) which leverages previously tagged posts providing appropriate suggestions for new content. Marinho [9] investigates the user-based collaborative approach for tag recommendation. The main outcome of this work shows that users with a similar tag vocabulary tend to tag alike, since the method seems to produce good results when applied on the user-tag matrix. The problem of tag recommendation through graph-based approaches has been firstly addressed by J¨aschke et al. in [7]. The key idea behind their FolkRank algorithm is that a resource which is tagged by important tags from important users becomes important itself. So, they build a graph whose nodes mutually reinforces themselves by spreading their weights. They compared some recommendation techniques including collaborative filtering, PageRank and FolkRank, showing that the FolkRank algorithm outperforms other approaches. Furthermore, Schmitz et al. [14] proposed association rule mining as a technique that might be useful in the tag recommendation process. In literature we can find also other methods (called hy-
brid), which try to integrate two of more sources of knowl- users bookmarked the same resource. they will receive ge(mainly, content and collaborative ones) in order to the same suggestions since the folksonomies built from improve the quality of recommended tags milar resources are the same Heymann et. al [6 present a tag recommender that ex- loits at the same time social knowledge and textual sources We will try to overcome these draw backs, by proposing an They produce recommendations exploiting both the HTML approach firstly based on the anal of similar resources ca- source code(extracting anchor texts and page texts)and the pable also of leveraging the tags already selected by the user annotations of the community. The effectiveness of this ap roach is also confirmed by the use of a large dataset crawled op of the tag rank. Figure 1 shows the general architecture from del icio us for the experimental evaluation. of STaR. The recommendation process is performed in four Lipczak in [8 proposes a similar hybrid approach. Firstly, steps, each of which is handled by a separate component the system extracts tags from the title of the resource. Af- 3.1 Indexing of resources der to expand the sets of candidate tags with the tags that Given a collection of resources(corpus) with some textual usually co-occur with terms in the title. Finally, tags are metadata(such as the title of the resource, the authors, the filtered and re-ranked exploiting the informations stored in description, etc. ) STaR firstly invokes the Indexer module a so-calledpersonomy'", the set of the tags previously used in order to perform a preprocessing step on these data b exploiting Apache Lucene. Obviously, the kind of metadata Finally, in [16] the authors proposed a model based on to be indexed is strictly dependant on the nature of the both textual contents and tags associated with the resource resources. For example, supposing to recommend tags for They introduce the concept of conflated tags to indicate a set bookmarks, we could index the title of the web page and the ource. Modeling in this way the existing tag space they are entries, we could index the title of the publication and of related tags(like blog, blogs, ecc. )used to annotate a re- extended description provided by users, while for Bibt able to suggest various tags for a given bookmark exploiting abstract. Let u be the set of users andn the cardinality both user and document models of this set, the indexing procedure is repeated N+ l times we build an index for each user(Personal Inder)storing the 3. STAR: A SOCIAL TAG RECOMMENDER index for the whole community(Social Inder)storing the information on the resources she previously tagged and an SYSTEM information about all the tagged resources by merging the ollowing the definition introduced in [7], a folksonomy singles Personal Indexes can be described as a triple(U, R, T)where Following the definitions presented above, given a user E U we define Personalinder(u)as U is a set of users: PersonalInder(u)={r∈Rt∈T:TAs(u,r)=t(1) · T is a set of tags where TAS is the tag assignment function TAS: UX R-T We can also define a tag assignment function TAS: UX which assigns tags to a resource annotated by a given user R→T Socialinder represents the union of all the user personal in- So, a collaborative tagging system is a platform compose f users, resources and tags that allows users to freely assign Sociallnder=U PersonalInder(u. tags to resources, while the tag recommendation task for a given user uEU and a resource r E R can be described as 3.2 Retrieval of Similar resources Next, STaR can take into account users requests in or- erated from a ranked set of candidate tags from which the der to produce personalized tag recommendations for each top n elements are suggested to the user resource. First, every user has to provide some information ommender system, developed at the University of Bari. The page or its URL ce to be tagged, such as the title of the Web STaR (Social Tag Recommender)is a content-based tag rec- about the resou inceptive idea behind STaR is to improve the model imple- sociated on\ system can identify the user since she has mented in systems like Tag Assist [15 or AutoTag [111 Next. if the Although we agree that similar resources usually share already posted other resources, it exploits data about he similar tags, in our opinion Mishne's approach presents two (language, the tags she uses more, the number of tags sh important draw backs: sually uses to annotate resources, etc. )in order to refine the query to be submitted against both the Social and Per 1. the tag re-ranking formula simply performs a sum of sonal indexes stored in Lucene. We used as query the title the occurrences of each tag among all the folksonomies of the web page( for bookmarks)or the title of the publica hout considering the similarity with the resource to tion(for BibTeX entries). Obviously before submitting th tagged. In this way tags often used to annotat query we processed it by deleting not useful characters and resources with a low similarity level could be ranked punctuatio In order to improve the performances of the lucene Query- 2. the proposed model does not take into account the ing Engine we replaced the original Lucene Scoring function previoustaggingactivityperformedbyusersIftwohttp://lucene.apacheorg
brid), which try to integrate two of more sources of knowledge (mainly, content and collaborative ones) in order to improve the quality of recommended tags. Heymann et. al [6] present a tag recommender that exploits at the same time social knowledge and textual sources. They produce recommendations exploiting both the HTML source code (extracting anchor texts and page texts) and the annotations of the community. The effectiveness of this approach is also confirmed by the use of a large dataset crawled from del.icio.us for the experimental evaluation. Lipczak in [8] proposes a similar hybrid approach. Firstly, the system extracts tags from the title of the resource. Afterwards, it performs an analysis of co-occurrences, in order to expand the sets of candidate tags with the tags that usually co-occur with terms in the title. Finally, tags are filtered and re-ranked exploiting the informations stored in a so-called ”personomy”, the set of the tags previously used by the user. Finally, in [16] the authors proposed a model based on both textual contents and tags associated with the resource. They introduce the concept of conflated tags to indicate a set of related tags (like blog, blogs, ecc.) used to annotate a resource. Modeling in this way the existing tag space they are able to suggest various tags for a given bookmark exploiting both user and document models. 3. STAR: A SOCIAL TAG RECOMMENDER SYSTEM Following the definition introduced in [7], a folksonomy can be described as a triple (U, R, T ) where: • U is a set of users; • R is a set of resources; • T is a set of tags. We can also define a tag assignment function tas: U × R → T. So, a collaborative tagging system is a platform composed of users, resources and tags that allows users to freely assign tags to resources, while the tag recommendation task for a given user u ∈ U and a resource r ∈ R can be described as the generation of a set of tags tas(u, r) ⊆ T according to some relevance model. In our approach these tags are generated from a ranked set of candidate tags from which the top n elements are suggested to the user. STaR (Social Tag Recommender) is a content-based tag recommender system, developed at the University of Bari. The inceptive idea behind STaR is to improve the model implemented in systems like TagAssist [15] or AutoTag [11]. Although we agree that similar resources usually share similar tags, in our opinion Mishne’s approach presents two important drawbacks: 1. the tag re-ranking formula simply performs a sum of the occurrences of each tag among all the folksonomies, without considering the similarity with the resource to be tagged. In this way tags often used to annotate resources with a low similarity level could be ranked first; 2. the proposed model does not take into account the previous tagging activity performed by users. If two users bookmarked the same resource, they will receive the same suggestions since the folksonomies built from similar resources are the same. We will try to overcome these drawbacks, by proposing an approach firstly based on the analysis of similar resources capable also of leveraging the tags already selected by the user during her previous tagging activity, by putting them on the top of the tag rank. Figure 1 shows the general architecture of STaR. The recommendation process is performed in four steps, each of which is handled by a separate component. 3.1 Indexing of Resources Given a collection of resources (corpus) with some textual metadata (such as the title of the resource, the authors, the description, etc.), STaR firstly invokes the Indexer module in order to perform a preprocessing step on these data by exploiting Apache Lucene6 . Obviously, the kind of metadata to be indexed is strictly dependant on the nature of the resources. For example, supposing to recommend tags for bookmarks, we could index the title of the web page and the extended description provided by users, while for BibteX entries, we could index the title of the publication and the abstract. Let U be the set of users and N the cardinality of this set, the indexing procedure is repeated N + 1 times: we build an index for each user (Personal Index) storing the information on the resources she previously tagged and an index for the whole community (Social Index) storing the information about all the tagged resources by merging the singles Personal Indexes. Following the definitions presented above, given a user u ∈ U we define P ersonalIndex(u) as: P ersonalIndex(u) = {r ∈ R|∃t ∈ T : tas(u, r) = t} (1) where tas is the tag assignment function tas: U × R → T which assigns tags to a resource annotated by a given user. SocialIndex represents the union of all the user personal indexes: SocialIndex = [N i=1 P ersonalIndex(ui) (2) 3.2 Retrieval of Similar Resources Next, STaR can take into account users requests in order to produce personalized tag recommendations for each resource. First, every user has to provide some information about the resource to be tagged, such as the title of the Web page or its URL, in order to crawl the textual metadata associated on it. Next, if the system can identify the user since she has already posted other resources, it exploits data about her (language, the tags she uses more, the number of tags she usually uses to annotate resources, etc.) in order to refine the query to be submitted against both the Social and Personal indexes stored in Lucene. We used as query the title of the web page (for bookmarks) or the title of the publication (for BibTeX entries). Obviously before submitting the query we processed it by deleting not useful characters and punctuation. In order to improve the performances of the Lucene Querying Engine we replaced the original Lucene Scoring function 6 http://lucene.apache.org
Filter opus Recommendation Figure 1: Architecture of STaR with an Okapi BM25 implementation. BM25 is nowadays Similar Resources onsidered as one of the state-of-the art retrieval models by IR community [12 Let d be a corpus of documents, d E D, BM25 retur 中0.8 the top-k resources with the highest similarity value given resource r(tokenized as a set of terms ti.. tm), and is defined as follows: Gazzetta. it sim(r, d) 0.4 ∑=1k1(4-b)+b,+n*d(t) where nt represents the occurrences of the term document d, I is the ratio between the length of the r Sim lar Resources and the average length of resources in the corpus. and b are two parameters typically set to 2.0 and 0.75 respec- tively, and idf (ti) represents the inverse document frequenc 0.7 of the term t defined as follows Inde +df(t)+0.5 idf(ti)=log df(ti)+0.5 Figure 2: Retrieval of Similar Resources where n is the number of resources in the collection and df(ti)is the number of resources in which the term ti occurs Given user u E U and a resource r, Lucene returns the resources whose similarity with r is greater or equal tha PersonalInder, instead, returns another online newspaper a threshold B. To perform this task Lucene uses both the Tuttosport. com). The similarity score returned by Lucene Personallnder of the user u and the sociallnder. more for- has been normalized mally 3.3 Extraction of Candidate Tags The role of the Tag Extractor is to produce as output PRes(u,q)={r∈ Personalinder()lim(q,n)≥ the list of the so-called"candidate tags"(namely, the tags considered as 'relevant by the tag recommender). In this tep the system gets the most similar resources returned S_Res(q)={r∈ SocialInder sim(q,r)≥ by the apache Lucene engine and builds their folksonomies Figure 2 depicts an example of the retrieving step. In (namely, the tags they have been annotated with). Next, it this case the target resource is represented by Gazzetta. it produces the list of candidate tags by computing for each one of the most famous Italian sport newspaper. Lucene tag from the folksonomy a score obtained by weighting the queries the SocialInder and returns as the most similar re- similarity score returned by Lucene with the normalized oc currence of the tag. If the Tag Extractor also gets the list ources an online newspaper( Corrieredellosport it)and the the most similar resources from the user PersonalInder, it official web site of an Italian Football Club(Inter. it).The will produce two partial folksonomies that are merged, as- http://nlp.uned.es/jperezi/lucene-bm25/ signing a weight to each folksonomy in order to boost the
Figure 1: Architecture of STaR with an Okapi BM25 implementation7 . BM25 is nowadays considered as one of the state-of-the art retrieval models by the IR community [12]. Let D be a corpus of documents, d ∈ D, BM25 returns the top-k resources with the highest similarity value given a resource r (tokenized as a set of terms t1 . . . tm), and is defined as follows: sim(r, d) = Pm i=1 n r ti k1((1−b)+b∗l)+nr ti ∗ idf(ti) (3) where n r ti represents the occurrences of the term ti in the document d, l is the ratio between the length of the resource and the average length of resources in the corpus. Finally, k1 and b are two parameters typically set to 2.0 and 0.75 respectively, and idf(ti) represents the inverse document frequency of the term ti defined as follows: idf(ti) = log N + df(ti) + 0.5 df(ti) + 0.5 (4) where N is the number of resources in the collection and df(ti) is the number of resources in which the term ti occurs. Given user u ∈ U and a resource r, Lucene returns the resources whose similarity with r is greater or equal than a threshold β. To perform this task Lucene uses both the PersonalIndex of the user u and the SocialIndex. More formally: P Res(u, q) = {r ∈ P ersonalIndex(u)|sim(q, r) ≥ β} S Res(q) = {r ∈ SocialIndex|sim(q, r) ≥ β} Figure 2 depicts an example of the retrieving step. In this case the target resource is represented by Gazzetta.it, one of the most famous Italian sport newspaper. Lucene queries the SocialIndex and returns as the most similar resources an online newspaper (Corrieredellosport.it) and the official web site of an Italian Football Club (Inter.it). The 7 http://nlp.uned.es/ jperezi/Lucene-BM25/ Figure 2: Retrieval of Similar Resources PersonalIndex, instead, returns another online newspaper (Tuttosport.com). The similarity score returned by Lucene has been normalized. 3.3 Extraction of Candidate Tags The role of the Tag Extractor is to produce as output the list of the so-called ”candidate tags” (namely, the tags considered as ’relevant’ by the tag recommender). In this step the system gets the most similar resources returned by the Apache Lucene engine and builds their folksonomies (namely, the tags they have been annotated with). Next, it produces the list of candidate tags by computing for each tag from the folksonomy a score obtained by weighting the similarity score returned by Lucene with the normalized occurrence of the tag. If the Tag Extractor also gets the list of the most similar resources from the user PersonalIndex, it will produce two partial folksonomies that are merged, assigning a weight to each folksonomy in order to boost the
tags previously used by the user Formally, for each query q(namely, the resource to b Table 1: Results comparing the lucene original scor tagged), we can a set of tags to recommend by build- ing function with BM25 ing two sets: candTagsp and candTagss. These sets are Scoring R defined as follows bookmark25.2629.6727.29 bookmark25.6236.6230.15 candTagsp(u,)=tETt=TAs(u, r)ArE P-Res(u, q)I Original bibtex 14.06 21.45 16.99 BM bibtex13.72229117. (q)={t∈m|t In the same way we can compute the relevance of each tag with respect to the query q BM25 overall 16.45 26 0如2 Original overall 16.43 23.58 19 relp(t, u, g) 2reP-Res(u,g)nf sim(r, q) (5) In the example in Figure 3, setting a threshold y= 0.20 the system would suggest the tags sport and newspaper. re,(,y)= sRes nr* sim(r, q) 4. EXPERIMENTAL EVALUATION designed two different experimental sessions to evalu where nf is the number of occurrences of the tag t in the an- ate the performance of the tag recommender. In the first ses- notation for resource r and nt is the sum of the occurrences sion we performed a comparison between the original scoring of tag t among all similar resources. function of Lucene and a novel BM25 implementation, while Finally, the set of Candidate Tags can be defined as: the second was carried out to tune the system parameters. 4.1 Description of the dataset candTags(u, q)=candTagsp(u, g)UcandTags, (q)(7) We designed the experimental evaluation by exploiting a where for each tag t the global relevance can be defined as: dataset gathered from Bibsonomy. It contains 263, 00, book mark posts and 158, 924 Bib TeX entries submitted by 3,617 ifferent users. For each of the 235.328 different URLs and rel(t, )=a*relp(t, q)+(1-a)*rels(t, q) the 143, 050 different BibTeX entries were also provided some where a(PersonalTagWeight)and(1-a)(SocialTagWeight) textual metadata(such as the title of the resource, the de- are the weights of the personal and social tags respectively. scription, the abstract and so on) Figure 3 depicts the procedure performed by the Tag er- We evaluated STaR by comparing the real tags(namely, tractor: in this case we have a set of 4 Social Tags(Newspa the tags a user adopts to annotate an unseen resource)with per, Online, Football and Inter)and 3 Personal Tags(Sport the suggested ones. The accuracy was finally computed us- Newspaper and Tuttosport). These sets are then merged ing classical IR metrics, such as Precision, Recall and Fl- building the set of Candidate Tags. This set contains 6 tags Measure. Precision(Pr) is defined as the number of relevant since the tag newspaper appears both in social and personal recommended tags divided by the number of recommended ags. The system associates a score to each tag that indi- tags. Recall (Re)is defined as the number of relevant rec- cates its effectiveness for the target resource. Besides, the ommended tags divided by the total number of relevant tag scores for the Candidate Tags are weighted again according available. The Fl-measure is computed by the following for to Social Tag Weight(a)and Personal Tag Weight(1-a)val- mula ues(in the example, 0.3 and 0.7 respectively), in order to boost the tags already used by the user in the final tag Indeed, we can point out that the social tag "football F1=(2*Pr+R the same score of the personal tag tuttosport, although its Pr+ Re origin 4.2 Experimental Session 1 3.4 Tag recommendation Firstly, we tried to evaluate the predictive accuracy of Finally, the last step of the recommendation process is STar comparing difference performed by the Filter. It removes from the list of ca Lucene original one and the aforementioned BM25 imple didate tags the ones not matching specific conditions, such mentation). We performed the same steps previously de- as a threshold for the relevance score computed by the Tag scribed, retrieving the most similar items using the two men- Extractor. Obviously, the value for the threshold and the tioned similarity functions and cor g the tags suggested maximum number of tags to be recommend is strictly de- by the system in both cases. Results are presented in Table pendent from the training data. Formally, given a user u E U, a query g and a thresh- In general, there is an improvement by adopting BM25 old value y, the goal of the filtering component is to build with respect to the lucene original similarity function. We rec(u, q) defined as follows can note that BM25 improved the both the recall (+6,95% for booking 46% for BibTeXs entries) and the F1 measure(+ 2, 86% for bookmarks, +0, 17% for BibTeXs en- ec(u, q)=t E candTags(u, q)rel(t, 9)>31
tags previously used by the user. Formally, for each query q (namely, the resource to be tagged), we can define a set of tags to recommend by building two sets: candT agsp and candT agss. These sets are defined as follows: candT agsp(u, q) = {t ∈ T|t = T AS(u, r) ∧ r ∈ P Res(u, q)} candT agss(q) = {t ∈ T|t = T AS(u, r) ∧ r ∈ S Res(q) ∧ u ∈ U} In the same way we can compute the relevance of each tag with respect to the query q as: relp(t, u, q) = P r∈P Res(u,q) n t r ∗ sim(r, q) nt (5) rels(t, q) = P r∈S Res(q) n t r ∗ sim(r, q) nt (6) where n t r is the number of occurrences of the tag t in the annotation for resource r and n t is the sum of the occurrences of tag t among all similar resources. Finally, the set of Candidate Tags can be defined as: candT ags(u, q) = candT agsp(u, q) ∪ candT agss(q) (7) where for each tag t the global relevance can be defined as: rel(t, q) = α ∗ relp(t, q) + (1 − α) ∗ rels(t, q) (8) where α (PersonalTagWeight) and (1−α) (SocialTagWeight) are the weights of the personal and social tags respectively. Figure 3 depicts the procedure performed by the Tag Extractor: in this case we have a set of 4 Social Tags (Newspaper, Online, Football and Inter) and 3 Personal Tags (Sport, Newspaper and Tuttosport). These sets are then merged, building the set of Candidate Tags. This set contains 6 tags since the tag newspaper appears both in social and personal tags. The system associates a score to each tag that indicates its effectiveness for the target resource. Besides, the scores for the Candidate Tags are weighted again according to SocialTagWeight (α) and PersonalTagWeight (1−α) values (in the example, 0.3 and 0.7 respectively), in order to boost the tags already used by the user in the final tag rank. Indeed, we can point out that the social tag ‘football’ gets the same score of the personal tag ‘tuttosport’, although its original weight was twice. 3.4 Tag Recommendation Finally, the last step of the recommendation process is performed by the Filter. It removes from the list of candidate tags the ones not matching specific conditions, such as a threshold for the relevance score computed by the Tag Extractor. Obviously, the value for the threshold and the maximum number of tags to be recommend is strictly dependent from the training data. Formally, given a user u ∈ U, a query q and a threshold value γ, the goal of the filtering component is to build rec(u, q) defined as follows: rec(u, q) = {t ∈ candT ags(u, q)|rel(t, q) > γ} Table 1: Results comparing the Lucene original scoring function with BM25 Scoring Resource Pr Re F1 Original bookmark 25.26 29.67 27.29 BM25 bookmark 25.62 36.62 30.15 Original bibtex 14.06 21.45 16.99 BM25 bibtex 13.72 22.91 17.16 Original overall 16.43 23.58 19.37 BM25 overall 16.45 26.46 20.29 In the example in Figure 3, setting a threshold γ = 0.20, the system would suggest the tags sport and newspaper. 4. EXPERIMENTAL EVALUATION We designed two different experimental sessions to evaluate the performance of the tag recommender. In the first session we performed a comparison between the original scoring function of Lucene and a novel BM25 implementation, while the second was carried out to tune the system parameters. 4.1 Description of the dataset We designed the experimental evaluation by exploiting a dataset gathered from Bibsonomy. It contains 263,004 bookmark posts and 158,924 BibTeX entries submitted by 3,617 different users. For each of the 235,328 different URLs and the 143,050 different BibTeX entries were also provided some textual metadata (such as the title of the resource, the description, the abstract and so on). We evaluated STaR by comparing the real tags (namely, the tags a user adopts to annotate an unseen resource) with the suggested ones. The accuracy was finally computed using classical IR metrics, such as Precision, Recall and F1- Measure. Precision (Pr) is defined as the number of relevant recommended tags divided by the number of recommended tags. Recall (Re) is defined as the number of relevant recommended tags divided by the total number of relevant tags available. The F1-measure is computed by the following formula: F1 = (2 ∗ P r ∗ Re) P r + Re (9) 4.2 Experimental Session 1 Firstly, we tried to evaluate the predictive accuracy of STaR comparing difference scoring function (namely, the Lucene original one and the aforementioned BM25 implementation). We performed the same steps previously described, retrieving the most similar items using the two mentioned similarity functions and comparing the tags suggested by the system in both cases. Results are presented in Table 1. In general, there is an improvement by adopting BM25 with respect to the Lucene original similarity function. We can note that BM25 improved the both the recall (+ 6,95% for bookmarks, +1,46% for BibTeXs entries) and the F1 measure (+ 2,86% for bookmarks, +0,17% for BibTeXs entries)
Similar Resources Similarity Sodal Tags 个0.8 Newspaper botball 02404=064 0.4 Scores 0.7 Tuttosport uttosport Figure 3: Description of the process performed by the Tag Extractor 4.3 Experimental Session 2 each resource in the dataset there are many textual fields Next we designed a second experimental evaluation in or- such as title, abstract, description, extended description, etc. der to compare the predictive accuracy of STar with differ- In this case we used as query the title of the webpage( for ent combinations of system parameters. Namely bookmarks)and the title of the publication(for BibTeX en- tries s). the maximum number of similar documents retrieved The last parameter we need to tune is the threshold te deem a tag as relevant (). We performed some tests sug. the value of a for the PersonalTag Weight and Social- gesting both 4 and 5 tags and we decided to recommend only 4 tags since the fifth was usually noisy. We also fixed the threshold value between 0.20 and 0.25 the threshold y to establish whether a tag is relevant In order to carry out this experimental session we used the aforementioned dataset both as training and test set. We ex- which fields of the target resource use to compose the ecuted the test over 50 000 bookmarks and 50 000 BibTeXs For each resource randomly chosen from the dataset and for each combination of parameters, we executed the following the best scoring function between Lucene standard one and Okapi BM25 · query preparation; First, tuning the number of similar documents to retrieve from the PersonalInder and Socialinder is very important, Lucene retrieval function invocation since a value too high can introduce noise in the retrieval cess, while a value too low can exclude documents cor building of the set of Candidate Tags: taining relevant tags. By analyzing the results returned by some test queries, we decided to set this value between 5 aring the recommended tags with the real tags and 10, depending on the training data. iated by the user Next, we tried to estimate the values for PersonalTag omputing of Precision, Recall, and Fl-measure Weight(PTW) and the SocialTag Weight(STW). An higher weight for the Personal Tags means that in the recommenda- Results are presented in Table 2 and Table 3 tion process the systems will weigh more the tags previously used by the target user, while an higher value for the So- Analyzing the results(see Figure ? ) it emerges that the cial Tags will give more importance to the tags used by the approach we called user-based outperformed the other ones. community(namely, the whole folksonomy)on the target In this configuration we set PTw to 1.0 and STw to 0, so resource. These parameters are biased by the user practice: we suggest only the tags already used by the user in tagging if tags often used by the user are very different from those similar resources. No query was submitted against the So- used from the community, the Ptw should be higher than cialInder. The first remark we can make is that each user STW. We performed an empirical study since it is difficult to has her own mental model and her own vocabulary: she usu- define the user behavior at run time. We tested the syster ally prefers to tag resources with labels she already used setting the parameters with several combinations of values: Instead, getting tags from the SocialInder only(as proved 1 PTW=0.7 STW=0 by the results of the community-based approach) often in- ii) PTW=0.5 STW=0.5 troduces some noise in the recommendation process. The iii) PTW=0.3 STW=0. hybrid approaches outperformed the community-based one Another parameter that can influence the system perfor but their predictive accuracy is still worse when compared ance is the set of fields to use to compose the query. For with the user-based approach. Finally, all the approaches
Figure 3: Description of the process performed by the Tag Extractor 4.3 Experimental Session 2 Next we designed a second experimental evaluation in order to compare the predictive accuracy of STaR with different combinations of system parameters. Namely: • the maximum number of similar documents retrieved by Lucene; • the value of α for the PersonalTagWeight and SocialTagWeight parameters; • the threshold γ to establish whether a tag is relevant; • which fields of the target resource use to compose the query; • the best scoring function between Lucene standard one and Okapi BM25. First, tuning the number of similar documents to retrieve from the PersonalIndex and SocialIndex is very important, since a value too high can introduce noise in the retrieval process, while a value too low can exclude documents containing relevant tags. By analyzing the results returned by some test queries, we decided to set this value between 5 and 10, depending on the training data. Next, we tried to estimate the values for PersonalTagWeight (PTW) and the SocialTagWeight (STW). An higher weight for the Personal Tags means that in the recommendation process the systems will weigh more the tags previously used by the target user, while an higher value for the Social Tags will give more importance to the tags used by the community (namely, the whole folksonomy) on the target resource. These parameters are biased by the user practice: if tags often used by the user are very different from those used from the community, the PTW should be higher than STW. We performed an empirical study since it is difficult to define the user behavior at run time. We tested the system setting the parameters with several combinations of values: i) PTW = 0.7 STW = 0.3; ii) PTW = 0.5 STW = 0.5; iii) PTW = 0.3 STW = 0.7. Another parameter that can influence the system performance is the set of fields to use to compose the query. For each resource in the dataset there are many textual fields, such as title, abstract, description, extended description, etc. In this case we used as query the title of the webpage (for bookmarks) and the title of the publication (for BibTeX entries). The last parameter we need to tune is the threshold to deem a tag as relevant (γ).We performed some tests suggesting both 4 and 5 tags and we decided to recommend only 4 tags since the fifth was usually noisy. We also fixed the threshold value between 0.20 and 0.25. In order to carry out this experimental session we used the aforementioned dataset both as training and test set. We executed the test over 50, 000 bookmarks and 50, 000 BibTeXs. For each resource randomly chosen from the dataset and for each combination of parameters, we executed the following steps: • query preparation; • Lucene retrieval function invocation; • building of the set of Candidate Tags; • comparing the recommended tags with the real tags associated by the user; • computing of Precision, Recall, and F1-measure. Results are presented in Table 2 and Table 3. Analyzing the results (see Figure ??), it emerges that the approach we called user-based outperformed the other ones. In this configuration we set PTW to 1.0 and STW to 0, so we suggest only the tags already used by the user in tagging similar resources. No query was submitted against the SocialIndex. The first remark we can make is that each user has her own mental model and her own vocabulary: she usually prefers to tag resources with labels she already used. Instead, getting tags from the SocialIndex only (as proved by the results of the community-based approach) often introduces some noise in the recommendation process. The hybrid approaches outperformed the community-based one, but their predictive accuracy is still worse when compared with the user-based approach. Finally, all the approaches
Table 2: Predictive accuracy of STaR over 50, 000 not similar items, maybe exploiting structured data or do- Approach STW PTw Pr typical Information Retrieval problem (namely, polysemy, synonymy, etc. )we will try to establish if the integration Conr 1.0 0.023.9624.6024.28 of Word Sense Disambiguation tools or a semantic repre 1.032.1228.7230.33 sentation of documents could improve the performance of recommender. Another issue to analyze is the application Hybrid 0.70.324.9626.3025.61 of our methodology in different domains such as multimedia Hybrid 0.50.524.1025.1624.62 environment. In this field discovering similarity among item Hybrid 0723.8525.1225.08 just on the ground of textual content could be not sufficient 35.5810.4216.11 Finally, we will perform also some studies in the area of ag-based recommendation, investigating the integration of tag recommenders for recommendations tasks. since reach- ing more quickly the tag convergence could help to build better folksonomies and to produce. more accurate recom- Table 3: Predictive accuracy of STaR over 50, 000 mendations BibTeXs STW PTW Pr Fl 6. REFERENCES Comm. based1.00.034.4435.8935.15 1 P. Basile, M. de Gemmis, P Lops, G Semeraro, 1.044.7340.5342.53 M. Bux, C. Musto, and F. Narducci. FIRSt: a ontent-based Recommender System Integrating Tags Hybrid 0.3323138.5735.16 for Cultural Heritage Personalization. In P Nesi Hybrid 35 375534.76 K. Ng, and J. Delgado, editors, Proceedings of the 4th 0.7 37.46 International Conference on Automated Solutions for Baselin 42.0313.2320.13 Cross Media Content and Multi-channel Distribution (AXMEDIS 2008). Workshop Panels and Industrial Applications, Florence, Italy, Firenze University Pres pages 103-106, November 17-19, 2008 outperformed the Fl-measure of the baseline. We compute 2 C H. Brooks and N Montanez Improved annotation the baseline recommending for each resource only its most nd hierarchical popular tags. Obviously, for resources never tagged we could clustering. In www 06: Proceedings of the 15th not suggest anything ternational conference on World wide web, page This analysis substantially confirms the results we ob- 625-632. New York. NY. USA. 2006. ACM Press. ained from other studies performed in the area of the tag 3 C. Cattuto, C Schmitz, A Baldassarri, V.D. P. based recommendation [ 1] ervedio, V. Loreto, A. Hotho, M. Grahl, and G. Stumme. Network properties of folksonomies. Al 5. CONCLUSIONS AND FUTURE WORK Communications, 20(4): 245-262, December 2007 Collaborative Tagging Systems are powerful tools, since 4 S Golder and B. A Huberman. The Structure of they let users to organize the information in a way that per- Collaborative Tagging Systems. Journal of fectly fits their mental model. However, typical drawbacks of Information Science, 32(2): 198-208, 2006 collaborative tagging systems represent an hindrance, since 5 Maria Grineva, Maxim Grinev, and Dmitry Lizorkin the complete tag space is too noisy to be exploited for re Extracting key terms from noisy and multi-theme trieval and filtering task. So, systems that assist I th documents. In 18th International world wide Web task of tagging speeding up the tag convergence are more Conference, pages 651-661, April 2009 and more required. In this paper we presented STar, a so- 间P al tag recommender system. The idea behind our work Social tag prediction. In SIGIR 08: Proceedings of the was to discover similarity among resources in order to ex- 31st annual international ACM SIGIR conference on ploit communities and user tagging behavior. In this way Research and development in information retrieval, our recommender system was able to suggest tags for users pages 531-538, New York, NY, USA, 2008. ACM and items still not stored in the training set. The exper 7R. Jaschke, L. Marinho, A Hotho mental sessions showed that users tend to reuse their own L. Schmidt-Thieme, and G. Stumme. Tag tags to annotate similar resources, so this kind of recommen- recommendations in folksonomies. In Alexander dation model could benefit from the use of the user personal Hinneburg, editor, Workshop Proceedings of Lernen tags before extracting the social tags of the community(we Wissensentdeckung- Adaptivit?t(LWA 2007), pag called this approach user-based ) Next, we showed that the 13-20, September 2007 integration of a more effective scoring function(BM25)could ] adation for folksonomies also improve the overall accuracy of the system oriented towards individual users. In Proceedings of This approach has a main drawback, since it cannot sug- ECML PKDD Discovery Challenge(RSDCo8), pages est any tags when the set of similar items returned by 84-95.2008 Lucene is empty. So, we plan to extend the system in or 9 L. B Marinho and L. Schmidt-Thieme. Collaborative der to extract significant keywords from the textual content tag recommendations. pages 533-540. 2008 associated to a resource(title, description, etc. ) that has [10 A. Mathes Folksonomies ative classification
Table 2: Predictive accuracy of STaR over 50, 000 bookmarks Approach STW PTW Pr Re F1 Comm.-based 1.0 0.0 23.96 24.60 24.28 User-based 0.0 1.0 32.12 28.72 30.33 Hybrid 0.7 0.3 24.96 26.30 25.61 Hybrid 0.5 0.5 24.10 25.16 24.62 Hybrid 0.3 0.7 23.85 25.12 25.08 Baseline - - 35.58 10.42 16.11 Table 3: Predictive accuracy of STaR over 50, 000 BibTeXs Approach STW PTW Pr Re F1 Comm.-based 1.0 0.0 34.44 35.89 35.15 User-based 0.0 1.0 44.73 40.53 42.53 Hybrid 0.7 0.3 32.31 38.57 35.16 Hybrid 0.5 0.5 32.36 37.55 34.76 Hybrid 0.3 0.7 35.47 39.68 37.46 Baseline - - 42.03 13.23 20.13 outperformed the F1-measure of the baseline. We computed the baseline recommending for each resource only its most popular tags. Obviously, for resources never tagged we could not suggest anything. This analysis substantially confirms the results we obtained from other studies performed in the area of the tagbased recommendation [1]. 5. CONCLUSIONS AND FUTURE WORK Collaborative Tagging Systems are powerful tools, since they let users to organize the information in a way that perfectly fits their mental model. However, typical drawbacks of collaborative tagging systems represent an hindrance, since the complete tag space is too noisy to be exploited for retrieval and filtering task. So, systems that assist users in the task of tagging speeding up the tag convergence are more and more required. In this paper we presented STaR, a social tag recommender system. The idea behind our work was to discover similarity among resources in order to exploit communities and user tagging behavior. In this way our recommender system was able to suggest tags for users and items still not stored in the training set. The experimental sessions showed that users tend to reuse their own tags to annotate similar resources, so this kind of recommendation model could benefit from the use of the user personal tags before extracting the social tags of the community (we called this approach user-based). Next, we showed that the integration of a more effective scoring function (BM25) could also improve the overall accuracy of the system. This approach has a main drawback, since it cannot suggest any tags when the set of similar items returned by Lucene is empty. So, we plan to extend the system in order to extract significant keywords from the textual content associated to a resource (title, description, etc.) that has not similar items, maybe exploiting structured data or domain ontologies. Furthermore, since tags usually suffer of typical Information Retrieval problem (namely, polysemy, synonymy, etc.) we will try to establish if the integration of Word Sense Disambiguation tools or a semantic representation of documents could improve the performance of recommender. Another issue to analyze is the application of our methodology in different domains such as multimedia environment. In this field discovering similarity among items just on the ground of textual content could be not sufficient. Finally, we will perform also some studies in the area of tag-based recommendation, investigating the integration of tag recommenders for recommendations tasks, since reaching more quickly the tag convergence could help to build better folksonomies and to produce more accurate recommendations. 6. REFERENCES [1] P. Basile, M. de Gemmis, P. Lops, G. Semeraro, M. Bux, C. Musto, and F. Narducci. FIRSt: a Content-based Recommender System Integrating Tags for Cultural Heritage Personalization. In P. Nesi, K. Ng, and J. Delgado, editors, Proceedings of the 4th International Conference on Automated Solutions for Cross Media Content and Multi-channel Distribution (AXMEDIS 2008) - Workshop Panels and Industrial Applications, Florence, Italy, Firenze University Press, pages 103–106, November 17-19, 2008. [2] C. H. Brooks and N. Montanez. Improved annotation of the blogosphere via autotagging and hierarchical clustering. In WWW ’06: Proceedings of the 15th international conference on World Wide Web, pages 625–632, New York, NY, USA, 2006. ACM Press. [3] C. Cattuto, C. Schmitz, A. Baldassarri, V. D. P. Servedio, V. Loreto, A. Hotho, M. Grahl, and G. Stumme. Network properties of folksonomies. AI Communications, 20(4):245–262, December 2007. [4] S. Golder and B. A. Huberman. The Structure of Collaborative Tagging Systems. Journal of Information Science, 32(2):198–208, 2006. [5] Maria Grineva, Maxim Grinev, and Dmitry Lizorkin. Extracting key terms from noisy and multi-theme documents. In 18th International World Wide Web Conference, pages 651–661, April 2009. [6] P. Heymann, D. Ramage, and H. Garcia-Molina. Social tag prediction. In SIGIR ’08: Proceedings of the 31st annual international ACM SIGIR conference on Research and development in information retrieval, pages 531–538, New York, NY, USA, 2008. ACM. [7] R. J¨aschke, L. Marinho, A. Hotho, L. Schmidt-Thieme, and G. Stumme. Tag recommendations in folksonomies. In Alexander Hinneburg, editor, Workshop Proceedings of Lernen - Wissensentdeckung - Adaptivit?t (LWA 2007), pages 13–20, September 2007. [8] M. Lipczak. Tag recommendation for folksonomies oriented towards individual users. In Proceedings of ECML PKDD Discovery Challenge (RSDC08), pages 84–95, 2008. [9] L. B. Marinho and L. Schmidt-Thieme. Collaborative tag recommendations. pages 533–540. 2008. [10] A. Mathes. Folksonomies - cooperative classification
and communication through shared metadata. WebsiteDecember2004.http://www.adammathes com/academic/computer-mediated-communication/ folksonomies html 11 G. Mishne. Autotag: a collaborative approach to automated tag assignment for weblog posts. In www 06: Proceedings of the 15th international conference on World wide Web, pages 953-954, New York, NY USA. 2006. ACM and M. Lau. Okapi at trec In Tert REtrieval ull [12S. E. Robertson, S. Walker, M. H. Beaulieu, A. Gull, Conference, pages 21-30, 1992 13 G. Salton. Automatic Tert Processing Addison-Wesley, 1989 14 C Schmitz, A. Hotho, R. Jaschke, and G. Stumme Mining association rules in folksonomies. I V. Batagelj, H.-H. Bock, A. Ferligoj, and A. Oiberna editors, Data Science and Classification(Proc. IFCS 2006 Conference), Studies in Classification, Data Analysis, and Knowledge Organization, pages 261-270, Berlin/Heidelberg, July 2006. Springer. Ljubljana [15 S Sood, S. Owsley, K. Hammond, and L. Birnbaum Tag Assist: Automatic Tag Suggestion for Blog Posts. Proceedings of the International Conference on Weblogs and Social Media(ICWSM 2007), 2007 16 M. Tatu, M. Srikanth, and T. D'Silva. Rsdc'08: Tag recommendations using bookmark content. In Proceedings of ECM PKDD Discovery Challenge ( RSDCO8), pages96-107,2008 17T. Vander Wal. Folksonomy coinage and definition. Website. Februar 2007 http://vanderwal.net/folksonomy.html 18 H. Wu, M. Zubair, and K. Maly. Harvesting social nowledge from folksonomies. In HYPERTEXT 06: Proceedings of the seventeenth conference on Hypertert and hypermedia, pages 111-114, New York, NY, USA 2006. ACM Press
and communication through shared metadata. Website, December 2004. http://www.adammathes. com/academic/computer-mediated-communication/ folksonomies.html. [11] G. Mishne. Autotag: a collaborative approach to automated tag assignment for weblog posts. In WWW ’06: Proceedings of the 15th international conference on World Wide Web, pages 953–954, New York, NY, USA, 2006. ACM. [12] S. E. Robertson, S. Walker, M. H. Beaulieu, A. Gull, and M. Lau. Okapi at trec. In Text REtrieval Conference, pages 21–30, 1992. [13] G. Salton. Automatic Text Processing. Addison-Wesley, 1989. [14] C. Schmitz, A. Hotho, R. J¨aschke, and G. Stumme. Mining association rules in folksonomies. In V. Batagelj, H.-H. Bock, A. Ferligoj, and A. Oiberna, ˝ editors, Data Science and Classification (Proc. IFCS 2006 Conference), Studies in Classification, Data Analysis, and Knowledge Organization, pages 261–270, Berlin/Heidelberg, July 2006. Springer. Ljubljana. [15] S. Sood, S. Owsley, K. Hammond, and L. Birnbaum. TagAssist: Automatic Tag Suggestion for Blog Posts. In Proceedings of the International Conference on Weblogs and Social Media (ICWSM 2007), 2007. [16] M. Tatu, M. Srikanth, and T. D’Silva. Rsdc’08: Tag recommendations using bookmark content. In Proceedings of ECML PKDD Discovery Challenge (RSDC08), pages 96–107, 2008. [17] T. Vander Wal. Folksonomy coinage and definition. Website, Februar 2007. http://vanderwal.net/folksonomy.html. [18] H. Wu, M. Zubair, and K. Maly. Harvesting social knowledge from folksonomies. In HYPERTEXT ’06: Proceedings of the seventeenth conference on Hypertext and hypermedia, pages 111–114, New York, NY, USA, 2006. ACM Press
