8 89 of users from user-generated tags. Analogous to our d a set of patterns, i.e., topics of inter- scovered provic systems. modeling description mendations of our approa sions are pre 2. Related work captur thus to build be ploit various aspects of user-g magnola et al. (2008)p enriched user model by analy meaning of tags. By applying system, namely icITY, th ndation navigating and tag e filteri users. Li et al. (2008) hat predict users'can be exploited to recommender systems. To build an individual user model, we connect tags and ratings as a way to infer a user’s topics of interest, in which each topic is composed of tags. In addition, to provide the model with more diversity, valuable topics in terms of both likes and dislikes are enriched in collaboration with other similar users. For recommending items relevant to user needs, we seamlessly incorporate collaborative characteristics into a content-based filtering approach so that recommender systems exploit the benefits of each, and thus alleviate the cold start problem and the overspecialization issue. The cold start problem describes a new user joins a recommender system and has presented few opinions (e.g., rating and tagging). With these situations, the system is generally unable to make high quality recommendations (Schein, Popescul, Ungar, & Pennock, 2002). The cold start users should be encouraged to continuously provide their opinions because they do not have enough historical information. However, inaccurate recommendations from a dearth of reliable information on the users lead them to undermine the credibility of the system, and thus may cause their deviation from the system. Considering this point, a differentiated strategy of building the user model is necessary to make compelling recommendations for those users. On the other hand, overspecialization refers to situations where a user model exclusively relies on tags which are used by a user or labeled in his/her rated items. With those cases, it is hard to recommend novel items different from anything the user has previously rated (Adomavicius & Tuzhilin, 2005). The main contributions of this paper toward user modeling in recommender systems can be summarized as follows: (i) We propose a new method of building a collaborative user model by leveraging user-generated tags. And we present a detailed method of topic-driven enrichments in collaboration with similar users that makes an individual model abundant. (ii) We present how the collaborative model can be applied to a recommender system. A locally weighted naïve Bayes approach is employed to recommend a ranked list of items relevant to users’ needs. We show how well our approach effectively works in terms of improving the recommendation quality and in dealing with cold start users. (iii) We present a new approach to identify two sets of similar neighbors by seamlessly combining rating information and tagging information. In particular, we separate out the neighborhood with respect to relevant tags from the neighborhood with respect to irrelevant tags. We demonstrate how two separated neighborhoods offer an advantage over a single set of the neighborhood. The rest of this paper is organized as follows: in next section we provide recent studies applying social tagging to recommender systems. In Section 3, we present a collaborative approach to user modeling for enriching valuable tags. We then provide a detailed description of how the proposed model is used for item recommendations in Section 4. In Section 5, we present the performance of our approach through experimental evaluations. Finally, conclusions are presented and future work is discussed in Section 6. 2. Related work The popularity of the usage of user-generated tags allows us to capture valuable information for understanding user interests and thus to build better user models. There are several studies that exploit various aspects of user-generated tags to user modeling. Carmagnola et al. (2008) proposed a new approach of constructing an enriched user model by analyzing users’ tagging behaviors and the meaning of tags. By applying the user model to a cultural heritage system, namely iCITY, the system supports personalized ways of navigating and tagging content, as well as recommend content to users. Li et al. (2008) presented a method of discovering common interests of users from user-generated tags. Analogous to our study, the authors discovered a set of patterns, i.e., topics of interest, that frequently co-occurred in tags added by users, and thus clustered users and resources according to each topic discovered. A similar approach is presented by Au Yeung, Cibbins, and Shadbolt (2009) as well. The authors constructed a user’s model in the form of a set of frequent tag patterns representing multiple interests of the user. The concept of discovering frequent tag patterns and subsequently grouping users according to tags in the patterns is close to our work. However, differing from these two studies, our work identifies and enriched further tag patterns of value to each user in collaboration with other similar users. More recently, in Wang, Clements, Wang, De Vries, and Reinders (2010), three types of personalization models in collaborative tagging systems are introduced according to user tasks: a collaborative tagging model, a collaborative browsing model, and a collaborative item search model. Guy, Zwerdling, Ronen, Carmel, and Uziel (2010) aggregated relationship information among users, tags, and items across various social media services within enterprise application environment. From those aggregated relationships, a user model is represented that contains a set of related users and a set of related tags. Wetzker et al. (2010) proposed a user-centric tag model mapping individual tag vocabularies, known as personomies, on the corresponding folksonomies. They also presented how the model can be applied to tag recommendations and tag-based social search, rather than item recommendations. In recent years, social tagging has also attracted considerable attention to recommender systems. In fact, many researchers have proposed a new application for recommender systems supporting the suggestion of suitable tags in annotating items. However, since our study focus on the collaborative nature of user modeling with social tagging for item recommendations, we mainly review studies dealing with item recommendations. Some early work in using tags for recommender systems is presented by Ji, Yeon, Kim, and Jo (2007). The authors first determine similarities between users with user-generated tags and subsequently identify the latent tags for each user by using a CF approach. Tso-Sutter, Marinho, and Thieme (2008) proposed a generic method that allows tags to be incorporated into standard CF algorithms by reducing three-dimensional correlations (i.e., user-item-tag) to three two-dimensional correlations and then applying a fusion method to re-associate these correlations. In Nakamoto et al. (2008), Reasonable Tag-based CF (RCF) is proposed assuming that tags generated by a user are synonymous with the reasons why the user liked an item. In RCF, tags are first clustered into topics by using an expectation-maximization (EM) algorithm. Afterwards, feature vectors of topics and items for each user, referred to as topic domain vectors, are created to find similar users and thus to recommend items in terms of topic domains. A Similar notion is previously described in De Gemmis, Lops, Semeraro, and Basile (2008) in which a user profile consists of two parts: profile_like and profile_dislike. The former contains features that help in finding relevant items whereas the latter helps in filtering out non-relevant items. A multivariate Poisson model for naïve Bayes is employed to estimate the posteriori probability and subsequently recommend a ranked list of items. There is the main difference between De Gemmis et al. (2008) and our work. Our work incorporates collaborative characteristics into a content-based recommendation that utilizes rating information and tagging information. On the other hand, De Gemmis et al. (2008) studied a typical content-based system that analyzes textual descriptions, such as summaries, reviews, and abstracts, in addition to ratings and user-generated tags. More recently, Siersdorfer and Sizov (2009) represented social tagging in the form of a vector space model that could apply to existing recommendation methods, i.e., content-based filtering and collaborative filtering. Sen and Vig (2009) introduced Tagommenders that predict users’ H.-N. Kim et al. / Expert Systems with Applications 38 (2011) 8488–8496 8489