Tagging activity has been recently identi ed as a potential source of knowledge about personal interests, preferences, goals, and other attributes known from user models. Tags themselves can be therefore used for nding personalized recommendations of items. In this paper, we present a tag-based recommender system which suggests similar Web pages based on the similarity of their tags from a Web 2.0 tagging application. The proposed approach extends the basic similarity calculus with external factors such as tag popularity, tag representativeness and the a nity between user and tag. In order to study and evaluate the recommender system, we have conducted an experiment involving 38 people from 12 countries using data from Del.icio.us, a social bookmarking web system on which users can share their personal bookmarks.
Collaborative tagging systems have become increasingly popular for sharing and
organizing Web resources, leading to a huge amount of user generated metadata.
Tags in social bookmarking systems such as del.ici.ous 1 are usually assigned to
conceptualize, categorize, or sharing a resource on the Web so that users can
be reminded of them later and nd their bookmarks in an easy way. Invariably,
tags represent some sort of a nity between user and a resource on the web. By
tagging, users label resources on the Internet freely and subjectively, based on
their sense of values [
The goal of this study is to analyze whether tags can be utilized to generate personalized recommendations. This assumption can be assessed by running an experiment whereby users expresses their satisfaction about the received recommendations. Based on this, we conducted an experiment involving 38 people from 12 countries using data from del.icio.us to evaluate the e ciency of the proposed approach and social aspects such as the purposes behind the tagging activity. The contribution of the paper is therefore: { The proposed recommendation approach based on similarity, tag representativeness, popularity, and a nity; and { Findings from the evaluation which show that the approach performs well in a non-controlled environment with people from di erent domains and intentions.
The paper is organized as follows: In Section 2 we discuss related work. In Section 3 we introduce the motivation of this work. Section 4 describes the factors of similarity that we will analyze. Section 5 presents the experiment and the achieved results. Section 6 addresses a discussion about the ndings from the experiment and Section 7 presents the conclusion and future works. 2
Tags have been recently studied in the context of recommender systems due to
various reasons. [
[
In this paper we look at, how multiple factors such as similarity, tag popularity, tag a nity to a user and tag representativeness can be used together to achieve recommendations. We also wanted to see the personalized recommendations in an open context with users of di erent background. 3
Tags in social bookmarking systems allow users to express their preference by
sharing their bookmarks. Tags are personalized piece of information which can
be utilized to identify common interests between users. Compared to traditional
collaborative rating [
Let us now look at a scenario which explains our approach and ideas behind it. According to Figure 1, the resources Google-Web and Wicket share tags ajax,programming andweb whereas resources Wicket and Swift only share tag web. Considering exclusively the similarity between tags, the resources GoogleWeb and Wicket have higher probability of being about the same content than Wicket and Swift. Based on this fact, their authors should be noticed about the existence of similar resources around. Furthermore, the noti cation could be prioritized (or emphasized) if the similar tags correspond to the most frequent tags of the authors or they are very representative for the resource they describe. This scenario was presented for illustrating how tag similarity can be computed for means of personalization. Although similarities can be found when the tags are syntactically identical, a number of pessimist scenarios may take place and must be considered such as: resources which have similar tags but incorrect spelling - since tags are informal and free writing, no syntax control is assured. For instance, tags "programming" and "programing" looks the same except for the fact the second one is missing the letter m; resources whose tags are syntactically di erent but similar semantically - this is a case of synonymy and to overcome it some semantic assistance is needed either by use of domain ontologies or looking up for synonymies in dictionary. For instance, tags "work" and "labor" looks di erent but share the same meaning. The obstacle is that generic dictionary sometimes is not enough to provide the correct meaning of speci c terms in a given context ; and resources which share same tags with di erent meanings - this is the well known case of polysemy. For instance, the tag "windows" can be about the operational system or the house artifact. 4
In order to generate personalized recommendations, we propose an extension method for the calculation of basic similarity between tags. We combine the cosine similarity calculus with other factors such as tag popularity, tag representativeness and a nity user-tag with the purpose of reordering the original raking in recommendation and generate personalized ones.
We de ne the document score as:
n n Ds = X weight(T agi) X representativness(T agi), where n is the total numi=1 i=1 ber of existing tags in the repository.
We de ne the tag user a nity as: Af f inity(u;t) = cardfr 2 Documents j (u; t; r) 2 R; R U T Dg=cardft 2 T j (t; u) 2 Ru; Ru U T g, where t is a particular tag, u particular user, U is a set of users, D set of resources and T set of tags.
Finally, similarity is computed as: Similarity(Di;Dii) = [DsDi +DsDii cosine similarity(TDi ; TDii)] Af f inity(u;t), where Ds is the document score and T is set of tags of a particular document.
Informally, each one of the factors in the above formulas is calculated as
follows:
{ Cosine Similarity | Our tag similarity is a variant on the classical cosine
similarity familiar from text mining and information retrieval [
Further, we have set empirically that for one tag represent the user's preference, its frequency of use must be 70% closer to the most frequent tag of the user. In the case on which there are no tags to satisfy this condition, it is assumed the user does not have a clear preference. 5
In our evaluation, we opted to measure the degree of satisfaction of users about the received recommendations. The user's feedback will allow us to evaluate the quality of recommendations produced from our framework. Although, we recognize that precision and recall are metrics which could be used to evaluate the e ectiveness of the system, we believe that user's participation provides more precious feedback for means of personalization. In this sense, we invited users by sending a number of invites in various mailing lists from di erent natures, not only related to technology. We explained the purpose of the experiment and also we outlined easiness of the participation aim at attracting more users not related to technology. To our surprise, within less than 1 month 44 participants had accepted to participate voluntarily. In spite of 44 initial positive replies, only 38 participants joined until the end. Finally, we had 38 participants from 12 countries interested in many di erent subjects. Data for our experiment was collected from del.icio.us in November 2008 comprising 5542 tags and 1143 bookmarks. Methodology. We have created a del.icio.us user account for each participant on which he/she was invited to add at least 10 bookmarks with minimally 3 tags each as suggested. Each participant received the top 5 most similar recommendations to their bookmarks based on the tags assigned to them. Then the participants were asked to select which items of the recommended set matched to their bookmarks. As soon as the participants nished their contribution, the overall results were shared with the participants as well as the re ections and ndings. 5.1
Considering that the experiment took place in non-controlled environment (as del.ici.ous is) with diverse audience (people from technology, health, education, biology, etc), we did not expected 100% of acceptance of the recommendations. Some reasons for this are: i) diversity of culture and background - Since the participants are from many di erent countries and have distinct backgrounds, it increases signi cantly the disparity between tags i)Syntax of tags - As previously introduced, the tags assigned by the participants were not under any syntax control. Users could have written their tags in many di erent (and personalized) ways, for instance, the tag web2.0 can be also tagged as web20, web2 0 or web 20 and iii) di culty to identify user's preferences through the tags - if users bookmark web resources of di erent domains (e.g. sports, education, engineering), hardly any some tags will predominate over others, which increases the di culty of precisely indentifying user's main preferences.
Based on the reasons addressed previously, we consider the result as satisfactory if more than half of recommendations are accepted (or selected) by the participants. If 80% of the recommendations are accepted, we claim the results as excellent (and unexpected), on the other hand, bellow of 50%, we understand that the proposal has to be reviewed and improved with the ndings achieved from the experiment analysis. 5.2
The graph shows that at least one recommendation from each set
recommended was accepted. Moreover, it shows that 8 participants were satis ed with
the whole set of recommendations and 7 participants accepted only 1 item from
whole the set which was recommended. Due to the graph distribution, it is
possible to preliminary argue that the acceptance is well balanced. However, in order
to evaluate the overall results more properly, we stipulated a threshold by
calculating the arithmetic mean of the acceptance, which was 2.971. In the following,
we calculated the standard error of the mean in order to verify the variance of
the mean and consequently perform more concrete argumentation on top of the
results obtained. The standard error is given by se = psn , where s is the sample
standard deviation (i.e. the measure of the dispersion of the data set), n is the
size (number of observations) of the sample. More about the standard deviation
can be found at [
Figure 3 shows the histogram of the accepted recommendations. It shows the frequency of the accepted items against the amount of participants. Figure 3 shows that 22 participants (or 58% of all) accepted 3 or more recommendations (above the threshold) from the ve that was suggested. However, 16 (or 42% of all) participants accepted only 2 or 1 recommendations, below the stipulated threshold. Focusing only on the 16 participants who accepted between 2 or 1 items, 7 of them accepted only 1 recommendation from the whole set. This means that 7 participants together rejected 80% of the recommendations that were sent to them (i.e. 35 sent and 28 rejected). In order to investigate this particular inconvenience, we analyzed the tags assigned to these rejected items. We gured out that although the recommendations had been generated correctly (considering the tag syntax), most of them was really out of context and far from the user's interest. We turn out with some ndings: i) some tags did not show clearly any relatedness with the resource domain; ii) ambiguity problems, more particularly synonymy, when the tags share same syntax but are semantically di erent; and iii) impossibility of identifying user's preference due to the low number of tags of some users.
Figure 4 summarizes the overall results. As already pointed, the pie chart on the left shows that 58% of the participants received a set of recommendations above the stipulated threshold while 42% received a set of recommendations bellow of it. The pie chart on the right shows the overall rate of well succeeded recommendations in which 59% of the whole recommendations was accepted and 41% was rejected. In summary, the nal result cannot be considered excellent but satisfactory, since 58% of the overall set of recommendations was above the threshold and 59% of the overall recommendations was accepted. In spite of achieving satisfactory results, it is not possible to a rm the proposed solution is ready for large usage. Improvements to overcome the ambiguity problems are needed and further experiments must be performed again in order to provide more insights about the system evolution. On the other hand, the results indicate the research is on the right track. 6
From the evaluation, we realized that the proposed recommender approach
performs satisfactorily well even in a non-controlled environment with users from
di erent domains and backgrounds. Based on the results, we understand the
multifactor approach can be utilized to generate personalized recommendations.
However, it is quite important to discuss the problems found from the
unsuccessful recommendations. The ambiguity problems such as synonymy and polysemy
can be attenuated by using WordNet dictionary since it has been employed for
computing semantic similarity [
This paper introduced a tag-based recommender system which generate personalized recommendations. The e ciency of the system rely on cosine similarity calculus with additional factors such as tag popularity, tag representativeness and a nity between user and tag. A experiment involving 38 people from 12 countries using data from del.icio.us was conducted to evaluate the e ciency of the system for means of personalization.
The overall results showed that approximately 60% of the recommendations succeeded and the proposed recommender system requires improvements. As a future work, we propose to perform semantic similarity to overcome ambiguity problems (as mentioned in the pessimist scenarios) and investigate the purpose of the tags when they are assigned to a resource. Finally, comparisons with other approaches must be addressed since the current evaluation methodology only assesses user's satisfaction using the speci c algorithm. 8
The research leading to these results is part of the project "KiWi - Knowledge in a Wiki" and has received funding from the European Communitys Seventh Framework Programme (FP7/2007-2013) under grant agreement No. 211932.