Thanks to the continuous growth of collaborative platforms like YouTube, Flickr and Delicious, we are recently witnessing to a rapid evolution of web dynamics towards a more 'social' vision, called Web 2.0. In this context collaborative tagging systems are rapidly emerging as one of the most promising tools. However, as tags are handled in a simply syntactical way, collaborative tagging systems suffer of typical Information Retrieval (IR) problems like polysemy and synonymy: so, in order to reduce the impact of these drawbacks and to aid at the same time the so-called tag convergence, systems that assist the user in the task of tagging are required. In this paper we present a system, called STaR, that implements an IR-based approach for tag recommendation. Our approach, mainly based on the exploitation of a stateof-the-art IR-model called BM25, relies on two assumptions: firstly, if two or more resources share some common patterns (e.g. the same features in the textual description), we can exploit this information supposing that they could be annotated with similar tags. Furthermore, since each user has a typical manner to label resources, a tag recommender might exploit this information to weigh more the tags she already used to annotate similar resources. We also present an experimental evaluation, carried out using a large dataset gathered from Bibsonomy.
H.3.1 [Information Storage and Retrieval]: Content Analysis and Indexing: Indexing methods; H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval: Information filtering 1.
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. 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 analyze the dynamics of collaborative tagging systems. 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 hard to exploit for retrieval or recommendation tasks without performing any form of processing.
This problem is a hindrance to completely exploit the expressive power of folksonomies, so 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: we refer to them as tag recommenders.
This paper presents STaR, a tag recommender system
implementing an IR-based approach that relies on a
state-ofthe-art IR model called BM25. In this work, already
presented [
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.
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 [
AutoTag [
The problem of tag recommendation through graph-based
approaches has been firstly addressed by Ja¨schke et al. in [
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 [
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. 3.1
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 Lucene2. Obviously, the kind of metadata to be indexed is strictly dependent on the nature of the resources. 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 Personal Indexes. 3.2
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.
In order to improve the performances of the Lucene
Querying Engine we replaced the original Lucene Scoring function
with an Okapi BM25 implementation3. BM25 is nowadays
considered as one of the state-of-the art retrieval models by
the IR community [
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:
m r sim(r, d) = X nti r ∗ idf (ti) i=1 k1((1 − b) + b ∗ l) + nti where ntri 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 (1) (2) 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 a user 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.
For example, we suppose that the target resource is represented by Gazzetta.it, one of the most famous Italian sport newspaper. Lucene queries the SocialIndex and it could returns as the most similar resources an online newspaper (Corrieredellosport.it) and the official web site of an Italian
3http://nlp.uned.es/ jperezi/Lucene-BM25/ Football Club (Inter.it). The PersonalIndex, instead, could return another online newspaper (Tuttosport.com). 2, setting a threshold γ = 0.20, the system would suggest the tags sport and newspaper. 3.3
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.
Figure 2 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
Finally, the last step of the recommendation process is performed by the Filter. It removes from the list of candidate tags those not matching specific conditions, such as a threshold for the relevance score computed by the Tag Extractor. Obviously, the value of the threshold and the maximum number of tags to be recommended are strictly dependent from the training data. In the example in Figure 4.
The goal of experimental session was to tune the system parameters in order to obtain the best effectiveness of the tag recommender. We exploited a large dataset gathered from Bibsonomy. 4.1
The dataset used for the experimental evaluation 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. 4.2
Firstly, we tried to evaluate the influence of different Lucene scoring functions on the performance of STaR. We randomly chose 10,000 resources from the dataset and we compared the results returned exploiting two different scoring functions (the Lucene original one and the BM25) in order to find the best one. 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 a low improvement by adopting BM25 with respect to the Lucene original similarity function. We can note that BM25 improved the recall of bookmarks (+ 6,95%) and BibTeX entries (+1,46%).
Next, using the BM25 as scoring function, we tried 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 Social• the threshold γ to establish whether a tag is relevant; • which fields of the target resource use to compose the query.
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). A 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 a 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:
User-based
Hybrid Hybrid Baseline 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. Results are presented in Table 2 and Table 3.
Analyzing the results, 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
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 tag-based recommendation [
Nowadays, collaborative tagging systems are powerful tools but they are affected from some drawbacks since the complete tag space is too noisy to be exploited for retrieval and filtering tasks. In this paper we presented STaR, a social tag recommender system. The idea behind our work was to discover similarity among resources exploiting a state-of-the-art IR-model called BM25. 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).
This approach has a main drawback, since it cannot suggest any tags when the set of similar items returned by Lucene is empty. We are planning to extend the system in order to extract significant keywords from the textual content associated to a resource (title, description, etc.) that has no similar items, maybe exploiting structured data or domain ontologies.
Furthermore, since tags usually suffer of typical Information Retrieval problem (polysemy, etc.) we will try to establish whether the integration of Word Sense Disambiguation algorithms or a semantic representation of documents could improve the performance of the recommender.
Anyhow, our approach resulted promising compared with already existing and state of the art approaches for tag recommendation. Indeed, our work classified in 6th position in the final results of the ECML-PKDD 2009 Discovery Challenge (id: 29723)4