《电子商务 E-business》阅读文献：Collaborative user modeling with user-generated tags for social recommender systems

Expert Systems with Applications 38(2011)8488-8496 Contents lists available at Science Direct Expert Systems with Applications ELSEVIER journalhomepagewww.elsevier.com/locate/eswa Collaborative user modeling with user-generated tags for social recommender systems Heung-Nam Kim,, Abdulmajeed Alkhaldi, Abdulmotaleb El Saddik., Geun-Sik Jo KI b School of Computer and Information Engineering. Inha University, 253 Younghyun-dong. Nam-gu, Incheon 402-751. Republic of Korea College of Computer and Information Sciences, King Saud University, Riyadh, Saudi arabia ARTICLE INFO ABSTRACT With the popularity of social media services, the sheer amount of content is increasing exponentially on the Social Web that leads to attract considerable attention to recommender systems. Recommender sys tems provide users with recommendations of items suited to their needs To provide proper recommen Recommender dations to users, recommender systems require an accurate user model that can reflect a users ocial media characteristics, preferences and needs. In this study, by leveraging user-generated tags as preference indi cators, we propose a new collaborative approach to user modeling that can be exploited to recommender systems. Our approach first discovers relevant and irrelevant topics for users, and then dual user model with collaboration from other similar users. In order to evaluate the our model, we compare experimental results with a user model based on coll pproaches and a vector space model. The experimental results have shown the proposed model provides a better representation in user interests and achieves better recommendation results in terms of accuracy e 2011 Elsevier Ltd. All rights reserved. 1 Introduction best use of"word-of-mouth"recommendations(Breese, man,& Kadie, 1998: Resnick, lacovo, Suchak, Bergstorm Social media has been changing the way people find information, 1994). Although the field of CF research has a large number share knowledge and communicate with each other. This social mation filtering problems, generally a typical CF domain starts with phenomenon has transformed the masses, who were only informa- rating information that maps user-item pairs on a set of numerical models in CF can be represented by ratings given 101m山m5hm able is increasing exponentially with daily additions. In turn, it is tagging. Consequently, a user model can profit by those tags in addi ecoming increasingly more difficult for users to find the most tion to ratings. Therefore, in the last few years, a number of studies attractive content and users often struggle with a great challenge have tried to combine recommender systems with social tagging in in terms of information overload( Siersdorfer Sizov, 2009). a way that can be highly beneficial to both areas(Milicevic, Nanop- Recommender systems that have emerged in response to the oulos, Ivanovic, 2010). Such systems generate automated recom- challenge provide users with recommendations of items that they mendations just as traditional recommender systems, but retain the would like the most(Adomavicius Tuzhilin, 2005). To provide flexibility of tagging information(Sen, vig, Riedl, 2009). In point of roper recommendations to users, the systems require information combining CF with tagging, they can be regarded as a new type of that includes a users characteristics, preferences, and needs, typi- hybrid recommender systems that utilize human perceptive con- cally referred to as a User Model(Godoy Amandi, 2005). Therefore, tent contained in items. The reason is that a set of aggregated tags building an accurate model for users is crucial to the success of rec on an item is rich and compact enough to characterize and describe ommender systems One of the most successful technologies among the same main concepts of the item although tag usage of users de- recommender systems is Collaborative Filtering(CF)that makes the ends on a type of media items(.g articles, music, videos, and phe tos)they annotate( Bischoff, Firan, Nejdl, Paiu, 2008: Li, Guo, Zhao, 2008: Wetzker, Zimmermann, Bauckhage, Albayrak, 2010). orresponding author. Tel. +1 613 562 5800x6248: fax: +1 613 562 5664 In this study, we introduce a new method of building a user E-mail address: hnkimemcrlab ottawa ca(H-N. Kim). model that can represent a user's diverse preferences, and thus 0957-4174 front matter o 2011 Elsevier Ltd. All rights reserved oi:10.1016eswa201101.048

Collaborative user modeling with user-generated tags for social recommender systems Heung-Nam Kim a,⇑ , Abdulmajeed Alkhaldi a , Abdulmotaleb El Saddik a,c , Geun-Sik Jo b a School of Information Technology and Engineering, University of Ottawa, 800 King Edward, Ottawa, Ontario, Canada K1N 6N5 b School of Computer and Information Engineering, Inha University, 253 Younghyun-dong, Nam-gu, Incheon 402-751, Republic of Korea c College of Computer and Information Sciences, King Saud University, Riyadh, Saudi Arabia article info Keywords: User modeling Personalization Recommender systems Social tagging Social media filtering abstract With the popularity of social media services, the sheer amount of content is increasing exponentially on the Social Web that leads to attract considerable attention to recommender systems. Recommender sys￾tems provide users with recommendations of items suited to their needs. To provide proper recommen￾dations to users, recommender systems require an accurate user model that can reflect a user’s characteristics, preferences and needs. In this study, by leveraging user-generated tags as preference indi￾cators, we propose a new collaborative approach to user modeling that can be exploited to recommender systems. Our approach first discovers relevant and irrelevant topics for users, and then enriches an indi￾vidual user model with collaboration from other similar users. In order to evaluate the performance of our model, we compare experimental results with a user model based on collaborative filtering approaches and a vector space model. The experimental results have shown the proposed model provides a better representation in user interests and achieves better recommendation results in terms of accuracy and ranking. 2011 Elsevier Ltd. All rights reserved. 1. Introduction Social media has been changing the way people find information, share knowledge and communicate with each other. This social phenomenon has transformed the masses, who were only informa￾tion consumers via mass media, to be producers of information. However, as rich information is shared through social media sites, the huge amount of information that has not previously been avail￾able is increasing exponentially with daily additions. In turn, it is becoming increasingly more difficult for users to find the most attractive content and users often struggle with a great challenge in terms of information overload (Siersdorfer & Sizov, 2009). Recommender systems that have emerged in response to the challenge provide users with recommendations of items that they would like the most (Adomavicius & Tuzhilin, 2005). To provide proper recommendations to users, the systems require information that includes a user’s characteristics, preferences, and needs, typi￾cally referred to as a User Model (Godoy & Amandi, 2005). Therefore, building an accurate model for users is crucial to the success of rec￾ommender systems. One of the most successful technologies among recommender systems is Collaborative Filtering (CF) that makes the best use of ‘‘word-of-mouth’’ recommendations (Breese, Hecker￾man, & Kadie, 1998; Resnick, Iacovou, Suchak, Bergstorm, & Riedl, 1994). Although the field of CF research has a large number of infor￾mation filtering problems, generally a typical CF domain starts with rating information that maps user-item pairs on a set of numerical values. Thus, user models in CF can be represented by ratings given by users on a set of items. Besides ratings, a number of modern ser￾vices also allow the users to add tags to the items, known as social tagging. Consequently, a user model can profit by those tags in addi￾tion to ratings. Therefore, in the last few years, a number of studies have tried to combine recommender systems with social tagging in a way that can be highly beneficial to both areas (Milicevic, Nanop￾oulos, & Ivanovic, 2010). Such systems generate automated recom￾mendations just as traditional recommender systems, but retain the flexibility of tagging information (Sen, Vig, & Riedl, 2009). In point of combining CF with tagging, they can be regarded as a new type of hybrid recommender systems that utilize human perceptive con￾tent contained in items. The reason is that a set of aggregated tags on an item is rich and compact enough to characterize and describe the same main concepts of the item although tag usage of users de￾pends on a type of media items (e.g., articles, music, videos, and pho￾tos) they annotate (Bischoff, Firan, Nejdl, & Paiu, 2008; Li, Guo, & Zhao, 2008; Wetzker, Zimmermann, Bauckhage, & Albayrak, 2010). In this study, we introduce a new method of building a user model that can represent a user’s diverse preferences, and thus 0957-4174/$ - see front matter 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.01.048 ⇑ Corresponding author. Tel.: +1 613 562 5800x6248; fax: +1 613 562 5664. E-mail address: hnkim@mcrlab.uottawa.ca (H.-N. Kim). Expert Systems with Applications 38 (2011) 8488–8496 Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa

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 addi￾tion, 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 prob￾lem and the overspecialization issue. The cold start problem describes a new user joins a recommender system and has pre￾sented few opinions (e.g., rating and tagging). With these situa￾tions, 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 informa￾tion. However, inaccurate recommendations from a dearth of reli￾able 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 recom￾mendations 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 pro￾pose a new method of building a collaborative user model by lever￾aging 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 col￾laborative model can be applied to a recommender system. A lo￾cally 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 recom￾mendation 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 informa￾tion. 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 recom￾mendations in Section 4. In Section 5, we present the performance of our approach through experimental evaluations. Finally, conclu￾sions 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 ex￾ploit various aspects of user-generated tags to user modeling. Car￾magnola 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 inter￾est, 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 sub￾sequently 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 per￾sonalization models in collaborative tagging systems are intro￾duced 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) aggre￾gated relationship information among users, tags, and items across various social media services within enterprise application envi￾ronment. From those aggregated relationships, a user model is rep￾resented that contains a set of related users and a set of related tags. Wetzker et al. (2010) proposed a user-centric tag model map￾ping 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 stud￾ies 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 incorpo￾rated into standard CF algorithms by reducing three-dimensional correlations (i.e., user-item-tag) to three two-dimensional correla￾tions and then applying a fusion method to re-associate these cor￾relations. In Nakamoto et al. (2008), Reasonable Tag-based CF (RCF) is proposed assuming that tags generated by a user are synony￾mous with the reasons why the user liked an item. In RCF, tags are first clustered into topics by using an expectation-maximiza￾tion (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 prob￾ability 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 tex￾tual descriptions, such as summaries, reviews, and abstracts, in addition to ratings and user-generated tags. More recently, Siers￾dorfer 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

8490 H.-N. Kim et aL/ Expert Systems with Applications 38(2011)8488-8496 preference for items by inferring the users' preference for tags. As shown in Fig. 1. from the ratings of items that have previously The authors extended content-based recommendation methods rated by a user, we classify the items into two portions: a seto&? in a way that learned relationships between tags and items sitive(relevant)items and a set of negative (irrelevant )items. based on inferred tag preferences and item ratings. In Zhen, Li, rating of a user for a certain item is greater than or equal to the and Yeung(2009), TagiCoFi is proposed integrating tagging infor- average rating Ru of the user, we then call this item a positive item ation into a model-based CF, particularly using relation regular- for him/her and a negative item otherwise. More formally, the set ized matrix factorization. Liang. Xu, Li, Nayak, and Tao (2010) of positive and negative items for user u are denoted as pos(u)and studied a user-based and an item-based CF combined with Neg(u) respectively such that Pos(u)=iEIRui> Ru. Neg(u)=i weighted tags, respectively. The authors considered all possible IRus Ru, and Pos(u)n Neg(u)=0. relationships between users, items and tags (i.e, user-item, tag- item and user-item similarities). 3. 1.2. Calculating weights of tags Similar to our motivation, the salient concept behind the above- In our study, we associate a weight of tags with a users rating. mentioned studies is that social tagging can be highly beneficial to rather than the well-known tf-idf weight. The calculation of the enhance the quality of recommender systems. Although the above- weight for tags is straightforward. First, the vector of rati mentioned studies give reasonable promise of improving the per- each user is normalized as rl= 1, and subsequently, the formance, our study is different from earlier work. Because users ized values are used for weights in terms of a user for a often use personal and self-reference tags, mostly helpful for them- mally, the weight of tag t annotated in item i for user u selves, we look into what a tagged item is about and how much a as wu(t), is computed by the user. To this end, in our study, we seamlessly integrate ratings wui(t)=Rui >,, r 2 (1) as explicit user interests and tags as implicit user interests. We then discover frequent topics and tags existing in a users set of both relevant and non-relevant items In addition rather than clus- Because a tag may appear in multiple items with different tering users according to the topics, we carry out topic-driven weights that quantify the interest of the tag for a certain user enrichments of a user model with collaboration from other users we compute the mean weight of the tag in the set of positive items so that the enriched model can reflect diverse topics not only rel- and negative items, respectively evant to user interests. but also irrelevant to user needs -(0x.增西ox则(0(2 3. Collaborative user modeling where P (t) is the set of positive items rated by user u containing In this section, we describe our approach to building a user tag t and lu(t) is the set of negative items rated by user u contain- model that is derived from the users ratings, as well as user-gen- ing tag t. Finally, the global weight of tag t for user u, denoted as uut. erated tags. Before going into further detail, the notation and def- can be illustrated by the following equation itions required for understanding our approach are introduced. Let U-1, U2,..., u be the set of all users, I=11, i2, (p+pm3)/2,ift∈m(t),t∈=(t) set of all items, and T-t1, tz,... tin be the set of all tags annotated Au ift∈F(t),tl(t) in items. Let r be a user-item rating matrix in which Rui represents ift∈s(t),t≠(t) the rating of user uE U on item iE L An element Rui either exists as numerical ordinal scale or is unknown. We denote unknown rat- ings by g. The matrix R can be decomposed into row vectors: 3. .3 DIscoverng topIcs each row vector represents the ratings of a particular user on each user. Because ew t tag patterns from a set of items rated by R=FI,., Ful with Fu=[Ru, 1,..., Ru n]. for every u E U. Therefore ry user has different tastes on items, a d ems. We also denote a set of items rated by a certain user u et used for the mining process should also be selected individually Iu=ie lVi: Rui+g. As for social tagging, a three-dimensional for each user. In the data mining research literature, frequent pat- tems)is projected onto a two-dimensional one lh ssign tags to terns are typically defined as patterns that occur at least as fre- quently as a predetermined threshold commonly referred to as a nly take tag-item relationships into account, informing how many this paper, we minimum support(Han, Pei, Yin, 2004). In our study, frequent imes a tag has been annotated in an item. We transform a bag- patterns are a set of tags that appear frequently together in either model of social tagging into a set-model indicating which tags oc- a set of a users positive items Pos(u)or a set of a user's negative cur and do not occur in a particular item. Formally, let Q be a tag ems Neg(u) As observed by previous studies(Bischoff et al, 2008: Li et al., item binary matrix in which Qi is set to 1 if item i has been anno- 2008), user-generated tags can cover the tated at least I times or more with tag t and 0 otherwise. ncepts of an ite and provide a higher-level abstraction on content of an item. Fur- thermore, according to a critical characteristic of social tagging, also 3. 1. Personal known as folksonomy, vocabularies that emerge organically from he tags can be considered to describe a particular topic for the item. n general, ratings or a user for Items renect now relevant or In this sense, if multiple tags are frequently annotated together with tags of an item are concise to the users understanding, and hence, a particular topic described apture content of the item(Li et al, 2008) It is worth examining hand, a set of tags appear frequently together in a user's set of neg- how we can connect the user's ratings and the tags assigned by ative items, the user might dislike a topic containing the tags Such other users in respect to the item. set of tags, i.e., a frequent tag pattern, can be discovered using exis ng frequent-pattern mining methods such as apriori scale of 1-5), each user would have his/ her own rating behavior. ative items, frequent tag patterns can also be divided

preference for items by inferring the users’ preference for tags. The authors extended content-based recommendation methods in a way that learned relationships between tags and items based on inferred tag preferences and item ratings. In Zhen, Li, and Yeung (2009), TagiCoFi is proposed integrating tagging infor￾mation into a model-based CF, particularly using relation regular￾ized matrix factorization. Liang, Xu, Li, Nayak, and Tao (2010) studied a user-based and an item-based CF combined with weighted tags, respectively. The authors considered all possible relationships between users, items and tags (i.e., user-item, tag￾item and user-item similarities). Similar to our motivation, the salient concept behind the above￾mentioned studies is that social tagging can be highly beneficial to enhance the quality of recommender systems. Although the above￾mentioned studies give reasonable promise of improving the per￾formance, our study is different from earlier work. Because users often use personal and self-reference tags, mostly helpful for them￾selves, we look into what a tagged item is about and how much a user prefers the item, rather than capturing what tags are used by the user. To this end, in our study, we seamlessly integrate ratings as explicit user interests and tags as implicit user interests. We then discover frequent topics and tags existing in a user’s set of both relevant and non-relevant items. In addition, rather than clus￾tering users according to the topics, we carry out topic-driven enrichments of a user model with collaboration from other users so that the enriched model can reflect diverse topics not only rel￾evant to user interests, but also irrelevant to user needs. 3. Collaborative user modeling In this section, we describe our approach to building a user model that is derived from the user’s ratings, as well as user-gen￾erated tags. Before going into further detail, the notation and def￾initions required for understanding our approach are introduced. Let U = u1, u2, ... , u|U| be the set of all users, I=i1, i2, ... , i|I| be the set of all items, and T = t1, t2, ... , t|T| be the set of all tags annotated in items. Let R be a user-item rating matrix in which Ru,i represents the rating of user u e U on item i e I. An element Ru,i either exists as numerical ordinal scale or is unknown. We denote unknown rat￾ings by £. The matrix R can be decomposed into row vectors: R ¼ ½~r1; ... ;~rjUj T with~ru ¼ ½Ru;1; ... ; Ru;jIj, for every u e U. Therefore, each row vector represents the ratings of a particular user on items. We also denote a set of items rated by a certain user u as Iu ¼ i 2 Ij8i : Ru;i – £. As for social tagging, a three-dimensional space between users, items, and tags (i.e., users assign tags to items) is projected onto a two-dimensional one. In this paper, we only take tag-item relationships into account, informing how many times a tag has been annotated in an item. We transform a bag￾model of social tagging into a set-model indicating which tags oc￾cur and do not occur in a particular item. Formally, let Q be a tag￾item binary matrix in which Qt,i is set to 1 if item i has been anno￾tated at least l times or more with tag t and 0 otherwise. 3.1. Personal user model from tags In general, ratings of a user for items reflect how relevant or interesting the items are to him/her. In addition, user-generated tags of an item are concise to the users understanding, and hence, capture content of the item (Li et al., 2008). It is worth examining how we can connect the user’s ratings and the tags assigned by other users in respect to the item. 3.1.1. Positive and negative items Although the rating scale is fixed as numerical values (e.g., a scale of 1–5), each user would have his/her own rating behavior. As shown in Fig. 1, from the ratings of items that have previously rated by a user, we classify the items into two portions: a set of po￾sitive (relevant) items and a set of negative (irrelevant) items. If a rating of a user for a certain item is greater than or equal to the average rating Ru of the user, we then call this item a positive item for him/her and a negative item otherwise. More formally, the set of positive and negative items for user u are denoted as Pos(u) and Neg(u), respectively such that PosðuÞ ¼ i 2 IjRu;i P Ru, NegðuÞ ¼ i 2 IjRu;i : ð3Þ 3.1.3. Discovering topics We discover frequent tag patterns from a set of items rated by each user. Because every user has different tastes on items, a data￾set used for the mining process should also be selected individually for each user. In the data mining research literature, frequent pat￾terns are typically defined as patterns that occur at least as fre￾quently as a predetermined threshold, commonly referred to as a minimum support (Han, Pei, & Yin, 2004). In our study, frequent patterns are a set of tags that appear frequently together in either a set of a user’s positive items Pos(u) or a set of a user’s negative items Neg(u). As observed by previous studies (Bischoff et al., 2008; Li et al., 2008), user-generated tags can cover the main concepts of an item and provide a higher-level abstraction on content of an item. Fur￾thermore, according to a critical characteristic of social tagging, also known as folksonomy, vocabularies that emerge organically from the tags can be considered to describe a particular topic for the item. In this sense, if multiple tags are frequently annotated together with each other in positive items of a user, the user would be interested in a particular topic described by the co-occurred tags. On the other hand, a set of tags appear frequently together in a user’s set of neg￾ative items, the user might dislike a topic containing the tags. Such set of tags, i.e., a frequent tag pattern, can be discovered using exist￾ing frequent-pattern mining methods such as Apriori (Agrawal & Srikant, 1994) and FP-growth (Han et al., 2004). Since a set of items rated by a user is divided into a set of positive items and a set of neg￾ative items, frequent tag patterns can also be divided into two por- 8490 H.-N. Kim et al. / Expert Systems with Applications 38 (2011) 8488–8496

H-N. Kim et aL/Expert Systems with Applications 38(2011)8488-8496 Negative items of User u Weights of Tag Positive Items of User u ★★★★★ tem3★★★ ★★tem7 a13 Tags m大2 Oau Tags o. 2 Hm6大大★ ★过m9k o iagn oa A [ TagL Item10★★ Fig. 1. Classification of positive items and negative items for a user according to his/her item ratings. tions according to the sets used for the mining process. We regard encounters serious limitations for finding a set of neighbors. In patterns discovered from the positive items as implicitly relevant practice, even when users are very active, the result of rated items topics, and patterns discovered from the negative items as implicitly is only a small proportion of the total number of items. Accord- irrelevant topics, defined as follows ingly, it is often the case that a pair of users has nothing in com- mon, and hence the similarity cannot be computed (Sarwar, Definition 1(Relevant topics). Let pattern p=, tz,..., tp be a set Karypis, Konstan, Reidl, 2001). Even when the computation of of tags such that p c T A relevant topic is defined as tag pattern p similarity is possible, it may not be very reliable, because insuffi- for which support of pattern p(ie, the ratio of items in the set cient information is processed. To that end, recent studies have Pos(u)that contains pattern p, written as Supu()) is greater than or determined similarities between users by using user-generated equal to the minimum support for us Bu tags instead of user ratings (i et al, 2007; Liang et al, 2010: Mar kines et al., 2009: Nakamoto et al., 2008: Siersdorfer Sizov 2009: Zhen et al, 2009). In this paper, we also identify the best neighbors Definition 2(Irrelevant topics). Analogous to a relevant topic, a based on tags labeled in items. However, differing from the previ- irrelevant topic is defined as a tag pattern for which the ratio of ous work, rating information is embedded into the tags when we items in the set Neg(u) that contains the pattern is greater than compute the similarities between users, rather than frequency- or equal to the minimum support for user u, Ou based weights for the tags. Moreover, as two users could like some According to the downward closure property of frequent pat- topics in common but could have a difference with respect to dis- terns, if a tag pattern is a frequent pattern, then all subsets of that likes, neighbors in terms of relevant topics are maintained to be pattern is also frequent patterns(Agrawal Srikant, 1994). As separated from neighbors in terms of irrelevant ones pointed out in Liet al. (2008) if all subsets of a certain pattern have In order to find k similar neighbors, the cosine similarity, which the same support value, then they are always annotated in the same quantifies the similarity of a pair of vectors according to their an- items among positive items(or negative items)of a user. In this gle, is employed to measure the similarity values between a target ase, the specific pattern may be more valuable to the userthan each user and every other user. Because a model of each user consists of individual pattern derived from the specific one. Inspired by Li et both relevant topics and irrelevant topics, we compute two simi- same as one another. We will henceforth use the term'topic'to refer and v in terms of relevant topics is measured by eg users, u by ToR, and Tolu the set of relevant topics and irrelevant topics, simpos Creme mn) ae Xkr (4) respectively. We present a formal description of a model for user r--LL x zeng--pr u as follows: IMu=(ToRu, Tolu), which is referred to as an initial mod el for him/her. To facilitate the following sections, we further define where ITu and ITp refer to the set of tags in relevant topics of user itive tag and a set of positive tags for user u is denoted as ITpo. and u with respect to irrelevant topics can also be calculated dr e some notations here. a tag occurring within the set ToRu is called a u and pectively. Analogously, the similarity between user u Analogously, a tag occurring within the set Tolu is called a negati tag and a set of negative tags for user u is denoted as ITweg 3. 2. Identifying tag-based neighborhood where Tu and IT refer to the set of tags in irrelevant topics of Before enriching a model for a certain user, we first need to user u and v, respectively form neighborhood of the user. The main goal of neig nation is to identify a set of users similar to a particular m The similarity value between a pair of users is in the range [0, 1] and the higher a users value, the more similar he/she is to a target monly referred to as neighbors. a typical collaborati user. Finally, for a given user E U, particular k users with the high-

tions according to the sets used for the mining process. We regard patterns discovered from the positive items as implicitly relevant topics, and patterns discovered from the negative items as implicitly irrelevant topics, defined as follows: Definition 1 (Relevant topics). Let pattern p = t1, t2, ... , t|p| be a set of tags such that p # T. A relevant topic is defined as tag pattern p for which support of pattern p (i.e., the ratio of items in the set Pos(u) that contains pattern p, written as Supu(p)) is greater than or equal to the minimum support for user u, hu. Definition 2 (Irrelevant topics). Analogous to a relevant topic, a irrelevant topic is defined as a tag pattern for which the ratio of items in the set Neg(u) that contains the pattern is greater than or equal to the minimum support for user u, hu. According to the downward closure property of frequent pat￾terns, if a tag pattern is a frequent pattern, then all subsets of that pattern is also frequent patterns (Agrawal & Srikant, 1994). As pointed out in Li et al. (2008), if all subsets of a certain pattern have the same support value, then they are always annotated in the same items among positive items (or negative items) of a user. In this case, the specific pattern may be more valuable to the user than each individual pattern derived from the specific one. Inspired by Li et al. (2008), we prune such subset patterns whose support values are the same as one another. We will henceforth use the term ‘topic’ to refer to the frequent tag patterns. Formally, for a given user u, we denote by ToRu and ToIu the set of relevant topics and irrelevant topics, respectively. We present a formal description of a model for user u as follows: IMu = hToRu, ToIui, which is referred to as an initial mod￾el for him/her. To facilitate the following sections, we further define some notations here. A tag occurring within the set ToRu is called a positive tag and a set of positive tags for user u is denoted as ITpos u . Analogously, a tag occurring within the set ToIu is called a negative tag and a set of negative tags for user u is denoted as ITneg u . 3.2. Identifying tag-based neighborhood Before enriching a model for a certain user, we first need to form neighborhood of the user. The main goal of neighborhood for￾mation is to identify a set of users similar to a particular user, com￾monly referred to as neighbors. A typical collaborative filtering encounters serious limitations for finding a set of neighbors. In practice, even when users are very active, the result of rated items is only a small proportion of the total number of items. Accord￾ingly, it is often the case that a pair of users has nothing in com￾mon, and hence the similarity cannot be computed (Sarwar, Karypis, Konstan, & Reidl, 2001). Even when the computation of similarity is possible, it may not be very reliable, because insuffi- cient information is processed. To that end, recent studies have determined similarities between users by using user-generated tags instead of user ratings (Ji et al., 2007; Liang et al., 2010; Mar￾kines et al., 2009; Nakamoto et al., 2008; Siersdorfer & Sizov, 2009; Zhen et al., 2009). In this paper, we also identify the best neighbors based on tags labeled in items. However, differing from the previ￾ous work, rating information is embedded into the tags when we compute the similarities between users, rather than frequency￾based weights for the tags. Moreover, as two users could like some topics in common but could have a difference with respect to dis￾likes, neighbors in terms of relevant topics are maintained to be separated from neighbors in terms of irrelevant ones. In order to find k similar neighbors, the cosine similarity, which quantifies the similarity of a pair of vectors according to their an￾gle, is employed to measure the similarity values between a target user and every other user. Because a model of each user consists of both relevant topics and irrelevant topics, we compute two simi￾larity values. Formally, the similarity between a pair of users, u and v in terms of relevant topics is measured by Eq. (4) simpos u;v ¼ P t2ðITpos u \ITpos v Þlpos u;v lpos u;t ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P t2ITpos u lpos2 u;t q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P t2ITpos u lpos2 v;t q ð4Þ where ITpos u and ITpos v refer to the set of tags in relevant topics of user u and v, respectively. Analogously, the similarity between user u and v with respect to irrelevant topics can also be calculated as: simneg u;v ¼ P t2ðITneg u \ITneg v Þlneg u;v lneg u;t ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P t2ITneg u lneg2 u;t q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P t2ITneg u lneg2 v;t q ð5Þ where ITneg u and ITneg v refer to the set of tags in irrelevant topics of user u and v, respectively. The similarity value between a pair of users is in the range [0, 1] and the higher a user’s value, the more similar he/she is to a target user. Finally, for a given user u e U, particular k users with the high￾Fig. 1. Classification of positive items and negative items for a user according to his/her item ratings. H.-N. Kim et al. / Expert Systems with Applications 38 (2011) 8488–8496 8491

H.-N. Kim et aL/ Expert Systems with Applications 38(2011)8488-8496 est similarity are identified as positive neighbors and negative neigh- only topic pz and p3 are used for enriching the model of user u if bors such that. they are not included in the set ToRu. Topics such as p2 and p3 are called enriched topics for user u, and tags such as t3 E P2 and Np (u)=arg max simp ta E p3 are called enriched tags for user u. Finally, a set of enriched topics is identified Nk(u)=arg max sim (6) neighbors of target user u. Note that some topics tha vEU\uN i froma other neighbor h such that simaLp> simp, for u+h If a particular 3.3. Enriching a user model topic is enriched many times from the neighbors, the target user is likely to be more interested in that topic. In an analogous fashion, Once we have identified the set of the nearest neighbors for a with respect to irrelevant topics, a set of enriched topics can be certain user u, his/her initial model IMu=(ToRu, Tol)described il determined from k negative neighbors Nk(u) of user u. Formally Section 3. is enriched rom the neighbors. The basic idea of EM,-(EToR. ETol where ETor, is the set of relevant topics en- to prefer similar tag patterns discovered within his her neighbors. ETolu is the set of irrelevant topics enriched from negative neigh- tive items of a user frequently contain tags "web2.0"and"semanticweb, "he/she may also be interested in rs such that ETolu n Tolu= o. For convenience we also define topics, such as"web2.0. ""semanticweb, "and"ontology, " that fre- any enriched tags occurring within either EToRu or ETolu as the quently appears in positive items of users similar to him/her. This ETo nlmp enriched tags from positive neighbors etp such that enrichment process is particularly effective to some users who do not contain topics sufficiently in their user model. ET such that ET nIT=0. Eventually, we represent user u As illustrated in Fig. 2, we elaborate on the general idea of to with the initial model IMu and the enriched model emu. We label driven enrichments in the foll the unified model the collaborative model for user u neighbor user V e npos(u) in descending order of similarity be- CMu(Mu. EMu) tween target user u and neighbors For each topic p; in ToRu, specific topics of pi in ToRy are identified. Given two topics Pi E ToRu and 4 Recommendation via probabilistic approach PjE ToRv, pi is said to a general topic of p; if and only if pi is a subset of pj. i.e.P. C Pr On the contrary, P, is called a specific topic of Pr For The final step is to generate recommendations such as a list of example, let pi=[t1 t2) be a relevant topic of user u such that top-N items that a user would like the most In our study, an item PiE ToRu, and ToR,=(p2, p3, p4. ps be the set of relevant topics of recommendation is treated classification problem that an and ps sftn that p2(, t2, t3), P3(t1,t2, ta). pa=(ti tz, ts, ts), item belongs to either a positive class Pos or a negative class of topic pi, they are said to the specific topics of pi whereas ps is adopt locally weighted naive Bayes( Frank, Hall, Pfahringer not the specific topic. In other words, pi is said to the general topic 2003) with the multinomial event model (McCallum Nigam, of topics p2, P3, and p4. Several specific topics occurring in the mod 1998)to recommend top-N items stochastically. To apply the el of neighbor v may be found. For efficient enrichment, we only naive Bayes classifier to the recommendation process of the target consider specific topics that have a higher support than that of a user u, for a given item i fIu, tags annotated in item i such that general topic. For example, assume that the support for Pl P2, P3, T-tETV: Q,. -l are used as feature variables. More formally, Sup(P2)=0.5. SupMp3)=0.47, and Supu(pa)=0.35). In this case, write ves model for computing and pa is 0.41, 0.5, 0.47, and 0.35, respectively (i.e. Supu(p1)=0.41. naive Ba sterior probability can Tagl 国图 neighbor v Tags ----- Enriched Topics Target User I Fig. 2. Enriching topics of a target user from similar neighbors

est similarity are identified as positive neighbors and negative neigh￾bors such that: Npos k ðuÞ ¼ arg max k v2Unfug simpos u;v ; Nneg k ðuÞ ¼ arg max k v2Unfug simneg u;v ð6Þ 3.3. Enriching a user model Once we have identified the set of the nearest neighbors for a certain user u, his/her initial model IMu = hToRu, ToIui described in Section 3.1 is enriched from the neighbors. The basic idea of enriching the model starts from assuming that the user is likely to prefer similar tag patterns discovered within his/her neighbors. For example, if positive items of a user frequently contain tags ‘‘web2.0’’ and ‘‘semanticweb,’’ he/she may also be interested in topics, such as ‘‘web2.0,’’ ‘‘semanticweb,’’ and ‘‘ontology,’’ that fre￾quently appears in positive items of users similar to him/her. This enrichment process is particularly effective to some users who do not contain topics sufficiently in their user model. As illustrated in Fig. 2, we elaborate on the general idea of topic￾driven enrichments in the following example: Firstly, we choose neighbor user _ 2 Npos k ðuÞ in descending order of similarity be￾tween target user u and neighbors. For each topic pi in ToRu, specific topics of pi in ToRv are identified. Given two topics pi e ToRu and pj e ToRv, pi is said to a general topic of pj if and only if pi is a subset of pj, i.e., pi  pj. On the contrary, pj is called a specific topic of pi. For example, let p1 = {t1, t2} be a relevant topic of user u such that p1 e ToRu, and ToRv = {p2, p3, p4, p5} be the set of relevant topics of user v such that p2 = {t1, t2, t3}, p3 = {t1, t2, t4}, p4 = {t1, t2, t3, t5}, and p5 = {t3, t4}. Since topic p2, p3, and p4 contain the entire tags of topic p1, they are said to the specific topics of p1 whereas p5 is not the specific topic. In other words, p1 is said to the general topic of topics p2, p3, and p4. Several specific topics occurring in the mod￾el of neighbor v may be found. For efficient enrichment, we only consider specific topics that have a higher support than that of a general topic. For example, assume that the support for p1, p2, p3, and p4 is 0.41, 0.5, 0.47, and 0.35, respectively (i.e., Supu(p1) = 0.41, Supv(p2) = 0.5, Supv(p3) = 0.47, and Supv(p4) = 0.35). In this case, only topic p2 and p3 are used for enriching the model of user u if they are not included in the set ToRu. Topics such as p2 and p3 are called enriched topics for user u, and tags such as t3 e p2 and t4 e p3 are called enriched tags for user u. Finally, a set of enriched topics is identified from k positive neighbors of target user u. Note that some topics that were previ￾ously enriched by a certain neighbor v are also identified from an￾other neighbor h such that simpos u;v P simpos u;h , for v – h. If a particular topic is enriched many times from the neighbors, the target user is likely to be more interested in that topic. In an analogous fashion, with respect to irrelevant topics, a set of enriched topics can be determined from k negative neighbors Nneg k ðuÞ of user u. Formally, the enriched model, for a given user u, is defined as a tuple EMu = hEToRu, EToIui where EToRu is the set of relevant topics en￾riched from positive neighbors such that EToRu \ ToRu ¼ £, and EToIu is the set of irrelevant topics enriched from negative neigh￾bors such that EToIu \ ToIu ¼ £. For convenience, we also define any enriched tags occurring within either EToRu or EToIu as the sets: a set of enriched tags from positive neighbors ETpos u such that ETpos u \ ITpos u ¼ £, and a set of enriched tags from negative neighbors ETneg u such that ETneg u \ ITneg u ¼ £. Eventually, we represent user u with the initial model IMu and the enriched model EMu. We label the unified model the collaborative model for user u: CMu = hIMu, EMui. 4. Recommendation via probabilistic approach The final step is to generate recommendations such as a list of top-N items that a user would like the most. In our study, an item recommendation is treated as a classification problem that an item belongs to either a positive class Pos or a negative class Neg depending on a collaborative model CMu. To this end, we adopt locally weighted naïve Bayes (Frank, Hall, & Pfahringer, 2003) with the multinomial event model (McCallum & Nigam, 1998) to recommend top-N items stochastically. To apply the naïve Bayes classifier to the recommendation process of the target user u, for a given item i R Iu, tags annotated in item i such that Ti = tj e T|" : Qj,i = 1 are used as feature variables. More formally, naïve Bayes model for computing a posterior probability can be written as: Fig. 2. Enriching topics of a target user from similar neighbors. 8492 H.-N. Kim et al. / Expert Systems with Applications 38 (2011) 8488–8496

P(Posl) P(Pos)TTP( Pos), P(Negli_(Ne,&apert Systems with Applications 38(2011)8488-8496 H-N. Kim er aL/ 8493 the movies as many as possible For minimizing noise tags and pro- P() Po PIG INeg) viding users'convenience, candidate tags were suggested to the ticipants by utilizing plot keywords and genres of each movie presented by IMDb. It should be noticed that our study is more inter- If we take the logarithm after dividing one by the other for a ested in capturing which tags are labeled in items rather than which log-likelihood ratio, then we can use the following equation tags are used by users. In total, we collected 10, 570 ratings on 592 items from 133 users (i.e, 133 rows and 592 columns of a user-item Zui=(In P(Pos)-In P(Neg)+2(InP(t Pos)-In P( Neg)( 8) matrix R). As for tagging data, we considered tags that were added to a particular movie by at least three more users because some tags were meaningless to other people except for users who create the where the prior probability for the class Pos and Neg are computed tags. We collected 63, 325 non-zero entries(ie, tag assignments on as P(Pos)= Pos(u)I/uI and P(Neg)=INeg(u)I/uI. respectively. In items)of a tag-item binary matrix Q after pruning some noise tags. addition, locally weighted conditional probabilities P(t lPos)and The experiments have been designed to answer the following PtINeg)with Laplacian smoothing are estimated as follow 1+)· P(tlPos)=V.+ EtenmuETo Ot Jut Is identifying two separate sets of neighbors effective for enriching user models? 1+J Is the enrichment of user models effective for improving accu- P(tNeg) u+∑ cte(r n,Or racy and ranking? Is the quality of recommendations based on the enriched model where fu and uy are respectively the frequency of positive tag ty competitive against the existing approaches? nd negative tag t in CMu. Vu is the total number of distinct tags Is the proposed approach able to provide proper recommenda- in CMu and o is the weight of tag j for user u given by tions even if users rated few items? ur,ift∈M To evaluate the performance of the recommendations, we di- ={ Hu jr if t与∈EMn与∈M (10) vided the dataset into a training set and a test set. For each user u in the dataset, we randomly selected 10 positive items, and subse- where user vis the most similar user who contributes to enrich tag! which the user previously rated were used as the training set to user u. This allows us to put larger weights on more relevant tags When building users' initial model from the training set, we set Finally a set of n ordered items that obtain the higher log-like- the minimum support O, for each user to 0. (202). To ensure that lihood ratios is recommended to user u such that. for each user, the experiments were repeated five times with TopN(u=arg maxzui ferent the training/test set. Therefore, the result values reported in the experiment sections are the averages over all five runs. As discussed before, recommender systems relying exclusively on a users rated items and his her tags may only recommend 5.2. Evaluation metrics items highly related to content the user has previously rated. It is hard to recommend novel items that are different from anything Error metrics such as Root Mean Square Error(RMSE)and Mean the user has previously rated, i.e., overspecialization probler Absolute Error(MAE), which are commonly used metrics for eval (Adomavicius& Tuzhilin, 2005). In our model, by utilizing enriched uating prediction accuracy, do not directly measure the quality of topics from neighbors similar to a particular user, items containing top-N recommendations( Cremonesi, Koren, Turrin, 2010: Mcl- enriched tags of great value to the user rank higher in the recom- aughin Herlocker, 2004). Therefore, we adopted two evaluation measures as follows: Precision at top N. In the context of top N recommendations, p 5. Experimental evaluation cision can judge how relevant a set of ranked recommendations is for users. precision measures the ratio of items in a list of recom- 5. 1. Experimental setup mendations that were also contained in the test set to number of items recommended(Herlocker, Konstan, Terveen, Riedl, 2004). Note that precision in recommender systems is generally an unde derived from rating information with tagging information, it was non-relevant items(McLaughin Herlocker, 2004). Precision at both tagging and rating information at the same time. Therefore, top N for a given user u in test set is giver we designed a Web-based interface that allows users to rate item P(uaN= Test(u)nTopN(u) with numerical values, as well as annotate them with tags. On hundred and thirty-three participants who are university students were invited to our experiment To gather sufficient rating info where Test(u)is the set of items rated by user u in the test data mation from the participants, we decided to select top box-office whereas TopN(u) is the set of top N items recommended to user u novies and top rated movies" in The Internet Movie Database Finally, the overall precision at top n for all users in the test set is (IMDb). After removing movies overlapped, 592 movies were ex- computed by averaging the personal precision tracted in total. Whenever the participants rated movies with 1-to- Ranking accuracy at top N. Although precision is an efficient 5 star scale, the participants were also encouraged to add tags to measure it does not consider an items rank within a list of top N items recommended. That is, an item recommended with top Nth ranking. Therefore, we also used the ranking accuracy metric

PðPosjiÞ ¼ PðPosÞ PðiÞ Y tj2Ti PðtjjPosÞ; PðNegjiÞ ¼ PðNegÞ PðiÞ Y tj2Ti PðtjjNegÞ ð7Þ If we take the logarithm after dividing one by the other for a log-likelihood ratio, then we can use the following equation: zu;i ¼ ðln PðPosÞ  ln PðNegÞÞ þX jTij j¼1 ln PðtjjPosÞ  ln PðtjjNegÞ  ð8Þ where the prior probability for the class Pos and Neg are computed as P(Pos)=|Pos(u)|/|Iu| and P(Neg)=|Neg(u)|/|Iu|, respectively. In addition, locally weighted conditional probabilities P(tj|Pos) and P(tj|Neg) with Laplacian smoothing are estimated as follows: PðtjjPosÞ ¼ 1 þ xj  f pos u;j Vu þ P t2ðITpos u [ETpos u Þxt  f pos u;t PðtjjNegÞ ¼ 1 þ xj  f neg u;j Vu þ P t2ðITneg u [ETneg u Þxt  f neg u;t ð9Þ where f pos u;j and f pos u;j are respectively the frequency of positive tag tj and negative tag tj in CMu. Vu is the total number of distinct tags in CMu and xj is the weight of tag j for user u given by: xj ¼ lu;j ; if tj 2 IMu lv;j ; if tj 2 EMu; tj 2 IMv 0; otherwise 8 >: ð10Þ where user v is the most similar user who contributes to enrich tag j to user u. This allows us to put larger weights on more relevant tags and smaller weights on irrelevant ones. Finally, a set of N ordered items that obtain the higher log-like￾lihood ratios is recommended to user u such that: TopNðuÞ ¼ arg max N i2InIu zu;i ð11Þ As discussed before, recommender systems relying exclusively on a user’s rated items and his/her tags may only recommend items highly related to content the user has previously rated. It is hard to recommend novel items that are different from anything the user has previously rated, i.e., overspecialization problem (Adomavicius & Tuzhilin, 2005). In our model, by utilizing enriched topics from neighbors similar to a particular user, items containing enriched tags of great value to the user rank higher in the recom￾mended set. 5. Experimental evaluation 5.1. Experimental setup As the focus of this study is to build an effective model of users derived from rating information with tagging information, it was difficult to find publicly available datasets that plentifully contain both tagging and rating information at the same time. Therefore, we designed a Web-based interface that allows users to rate items with numerical values, as well as annotate them with tags. One hundred and thirty-three participants who are university students were invited to our experiment. To gather sufficient rating infor￾mation from the participants, we decided to select top box-office movies1 and top rated movies2 in The Internet Movie Database (IMDb). After removing movies overlapped, 592 movies were ex￾tracted in total. Whenever the participants rated movies with 1-to- 5 star scale, the participants were also encouraged to add tags to the movies as many as possible. For minimizing noise tags and pro￾viding users’ convenience, candidate tags were suggested to the par￾ticipants by utilizing plot keywords and genres of each movie presented by IMDb. It should be noticed that our study is more inter￾ested in capturing which tags are labeled in items rather than which tags are used by users. In total, we collected 10,570 ratings on 592 items from 133 users (i.e., 133 rows and 592 columns of a user-item matrix R). As for tagging data, we considered tags that were added to a particular movie by at least three more users because some tags were meaningless to other people except for users who create the tags. We collected 63,325 non-zero entries (i.e., tag assignments on items) of a tag-item binary matrix Q after pruning some noise tags. The experiments have been designed to answer the following questions: – Is identifying two separate sets of neighbors effective for enriching user models? – Is the enrichment of user models effective for improving accu￾racy and ranking? – Is the quality of recommendations based on the enriched model competitive against the existing approaches? – Is the proposed approach able to provide proper recommenda￾tions even if users rated few items? To evaluate the performance of the recommendations, we di￾vided the dataset into a training set and a test set. For each user u in the dataset, we randomly selected 10 positive items, and subse￾quently used those as his/her test set. And the remaining items which the user previously rated were used as the training set. When building users’ initial model from the training set, we set the minimum support hu for each user to 0.2 (20%). To ensure that our results are not sensitive to the particular training/test portion￾ing for each user, the experiments were repeated five times with different the training/test set. Therefore, the result values reported in the experiment sections are the averages over all five runs. 5.2. Evaluation metrics Error metrics such as Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), which are commonly used metrics for eval￾uating prediction accuracy, do not directly measure the quality of top-N recommendations (Cremonesi, Koren, & Turrin, 2010; McL￾aughin & Herlocker, 2004). Therefore, we adopted two evaluation measures as follows: Precision at top N. In the context of top N recommendations, pre￾cision can judge how relevant a set of ranked recommendations is for users. Precision measures the ratio of items in a list of recom￾mendations that were also contained in the test set to number of items recommended (Herlocker, Konstan, Terveen, & Riedl, 2004). Note that precision in recommender systems is generally an under￾estimate of the true precision because we treat non-rated items as non-relevant items (McLaughin & Herlocker, 2004). Precision at top N for a given user u in test set is given: PðuÞ@N ¼ jTestðuÞ \ TopNðuÞj jTopNðuÞj ð12Þ where Test(u) is the set of items rated by user u in the test data whereas TopN(u) is the set of top N items recommended to user u. Finally, the overall precision at top N for all users in the test set is computed by averaging the personal precision. Ranking accuracy at top N. Although precision is an efficient measure, it does not consider an item’s rank within a list of top N items recommended. That is, an item recommended with top ranking is treated equally with an item that is recommended with Nth ranking. Therefore, we also used the ranking accuracy metric, 1 http://www.imdb.com/boxoffice/alltimegross. 2 http://www.imdb.com/chart/top. H.-N. Kim et al. / Expert Systems with Applications 38 (2011) 8488–8496 8493

8494 H.-N. Kim et aL/ Expert Systems with Applications 38(2011)8488-8496 which was introduced by Deshpande and Karypis(2004). The rank- the useful information has already been enriched well by more ng accuracy of user u at top N is defined as similar neighbors. In consideration of both precision and raking accuracy, the (13) neighborhood size for enriching the model of each user was set to 40 in subsequent experiments. where rank(i), 1 rank(< N, refers to the recommended rank of 5.4. Effect of model enrichment m i within the top N list of user u. That is, relevant items that ap- pear earlier in the top-N list are given more weight than later ones This section investigates the effect of the collaborative mod The higher the RK(u)oN value, the more accurately the algorithm CMu of each user in more detail, by comparing the results obtained ranks items to user u. Finally the overall raking accuracy for all by the initial model IMu of each user. users is computed by averaging the personal raking value in the tes Fig 3 shows the results of the exper t. The results demon data strate that the collaborative model enriched by neighbors en- hances the performance of the initial model. With respect to precision, CUM achieves 3.6% improvement More importantly, it 5.3. Experiments with neighborhood size is apparent that the collaborative model considerably improved The following experiment investigates the influence of neigh- larly important because users tend to be concerned with items bors on the enrichment of the user model. For this reason, different having higher ranks. The results confirm that the enriched tags numbers of neighbors k were used for model enrichment: 10, 20, and topics positively impact on the user model and consequently 30, 40, 50, and 60. In addition, we also adopted a standard similar- the collaborative model has advantages in terms of improving both ty measurement, which employed the cosine similarity using the the recommendation accuracy and its ranking. rating matrix R, during the neighborhood formation of each user In fact, ditterent measurements lead to a ditterent set of neighbors 5.5. Comparisons with other methods for a user, in turn, leading to different topics and tags enriched. Table 1 and Table 2 summarize the results according to differ- ent numbers of neighbors. The rows labeled"Pos/ Neg"show the re In this section, we present detailed experimental results in omparison with baseline methods. For comparison purposes, we and negative neighbors described in Section 3.2. And the rows la- tested the following baselines a cF based on a user-to-user sim- beledratingSim"show the results of the collaborative model en- similarity ITI( Deshpande& Karypis, 2004), and(ii)a tf-idf vector riched from rating-based standard neighbors. Comparing the results achieved by Pos/ Neg and ratingSim, the pace model using tags vSM. Our collaborative model CUM was enriched model which is derived from positive and negative neigh dations. As pointed out in the previous studies(cremonesi et d then compared with the baseline algorithms in top-N recomme bors separately is more accurate than the model! formed from the 2010: Mclaughin Herlocker, 2004), directly using predicted rat single neighborhood. For example, when the neighborhood size k 10, Pos/Neg achieves 3. 26% improvement over ratingSim in terms In fact, of P@10. With respect to RK@10, similar results are demonstrated. on predicted ratings. However, the performance on precision Nevertheless, we observed that the model based on rating Sim also remarkably low even though prediction accuracy was quite good seemed to work well. That may be caused by rating density in the Therefore, CF methods are designed to compute ranking scores matrix R. The density of the rating matrix we used was relatively high,and thus, we could identify reliable neighbors using rating sine similarity between two users(for UTU)or two items(for ITm)as information. Examining the value of each case in small sizes of the weight. We also conducted the parameter tuning experiment the neighborhood, the both cases provide a reasonably good per- with UTU and Ii beforehand in order to choose the best neighbor ormance in comparison with large sizes of the neighborhood. This hood size. Consequently, the size of UTU and ITI was set to 80 and result shows that the accuracy of our model is relatively insensitive 50, respectively. As for VSM, it was designed to build a user model to the value of k That is, once the number of nearest neighbors is from tags in his her positive and negative items. Simply, we sub- user is barely changed by any further increases in the number of vectors of the positive items. Afterwards an item was ranked using ranking values appeared In other words, the neighborhood with the calculated cosine similarity between tags annotated in the item a small size provides enough to enrich topics for each user because and the user model ■ UM CUM Precision at top 10 with respect to increasing k value. eighbors (k) 10 0.6 Pos/Neg 022350222022100.2210022470.2273 ratingsin0.1909020100.19090.19090.19700.1s 0 Ranking accuracy at top 10 with respect to increasing k value. eighbors(k) 10 20 40 Pos/Neg 0.82490.84910848408596084470856 081720.81770817208125082660.8122 Fig. 3. Comparison of P@10 and RKe10 obtained by the initial model and the

which was introduced by Deshpande and Karypis (2004). The rank￾ing accuracy of user u at top N is defined as: RKðuÞ@N ¼ X i2TestðuÞ\TopNðuÞ 1 rankðiÞ ð13Þ where rank(i), 1 6 rank(i) 6 N, refers to the recommended rank of item i within the top N list of user u. That is, relevant items that ap￾pear earlier in the top-N list are given more weight than later ones. The higher the RK(u)@N value, the more accurately the algorithm ranks items to user u. Finally, the overall raking accuracy for all users is computed by averaging the personal raking value in the test data. 5.3. Experiments with neighborhood size The following experiment investigates the influence of neigh￾bors on the enrichment of the user model. For this reason, different numbers of neighbors k were used for model enrichment: 10, 20, 30, 40, 50, and 60. In addition, we also adopted a standard similar￾ity measurement, which employed the cosine similarity using the rating matrix R, during the neighborhood formation of each user. In fact, different measurements lead to a different set of neighbors for a user, in turn, leading to different topics and tags enriched. Table 1 and Table 2 summarize the results according to differ￾ent numbers of neighbors. The rows labeled ‘‘Pos/Neg’’ show the re￾sults of the user model in collaboration with positive neighbors and negative neighbors described in Section 3.2. And the rows la￾beled ‘‘ratingSim’’ show the results of the collaborative model en￾riched from rating-based standard neighbors. Comparing the results achieved by Pos/Neg and ratingSim, the enriched model which is derived from positive and negative neigh￾bors separately is more accurate than the model formed from the single neighborhood. For example, when the neighborhood size k is 10, Pos/Neg achieves 3.26% improvement over ratingSim in terms of P@10. With respect to RK@10, similar results are demonstrated. Nevertheless, we observed that the model based on ratingSim also seemed to work well. That may be caused by rating density in the matrix R. The density of the rating matrix we used was relatively high, and thus, we could identify reliable neighbors using rating information. Examining the value of each case in small sizes of the neighborhood, the both cases provide a reasonably good per￾formance in comparison with large sizes of the neighborhood. This result shows that the accuracy of our model is relatively insensitive to the value of k. That is, once the number of nearest neighbors is relatively large enough, the rank of recommended items for each user is barely changed by any further increases in the number of neighbors even though the slight variation of the precision and ranking values appeared. In other words, the neighborhood with a small size provides enough to enrich topics for each user because the useful information has already been enriched well by more similar neighbors. In consideration of both precision and raking accuracy, the neighborhood size for enriching the model of each user was set to 40 in subsequent experiments. 5.4. Effect of model enrichment This section investigates the effect of the collaborative model CMu of each user in more detail, by comparing the results obtained by the initial model IMu of each user. Fig. 3 shows the results of the experiment. The results demon￾strate that the collaborative model enriched by neighbors en￾hances the performance of the initial model. With respect to precision, CUM achieves 3.6% improvement. More importantly, it is apparent that the collaborative model considerably improved the ranking values, compared to the initial model. This is particu￾larly important because users tend to be concerned with items having higher ranks. The results confirm that the enriched tags and topics positively impact on the user model and consequently the collaborative model has advantages in terms of improving both the recommendation accuracy and its ranking. 5.5. Comparisons with other methods In this section, we present detailed experimental results in comparison with baseline methods. For comparison purposes, we tested the following baselines: (i) a CF based on a user-to-user sim￾ilarity UTU (Breese et al., 1998), (ii) a CF based on an item-to-item similarity ITI (Deshpande & Karypis, 2004), and (iii) a tf-idf vector space model using tags VSM. Our collaborative model CUM was then compared with the baseline algorithms in top-N recommen￾dations. As pointed out in the previous studies (Cremonesi et al., 2010; McLaughin & Herlocker, 2004), directly using predicted rat￾ings as ranking score may not accurately recommend items. In fact, we generated top-N recommendations with CF approaches based on predicted ratings. However, the performance on precision is remarkably low even though prediction accuracy was quite good. Therefore, CF methods are designed to compute ranking scores achieved by non-normalized weighted sum of ratings using the co￾sine similarity between two users (for UTU) or two items (for ITI) as the weight. We also conducted the parameter tuning experiment with UTU and ITI beforehand in order to choose the best neighbor￾hood size. Consequently, the size of UTU and ITI was set to 80 and 50, respectively. As for VSM, it was designed to build a user model from tags in his/her positive and negative items. Simply, we sub￾tracted tag weight vectors of the negative items from tag weight vectors of the positive items. Afterwards an item was ranked using the calculated cosine similarity between tags annotated in the item and the user model. Table 1 Precision at top 10 with respect to increasing k value. Neighbors (k) 10 20 30 40 50 60 Pos/Neg 0.2235 0.222 0.2210 0.2210 0.2247 0.2273 ratingSim 0.1909 0.2010 0.1909 0.1909 0.1970 0.1919 Table 2 Ranking accuracy at top 10 with respect to increasing k value. Neighbors (k) 10 20 30 40 50 60 Pos/Neg 0.8249 0.8491 0.8484 0.8596 0.8447 0.8562 ratingSim 0.8172 0.8177 0.8172 0.8125 0.8266 0.8122 Fig. 3. Comparison of P@10 and RK@10 obtained by the initial model and the collaborative model. 8494 H.-N. Kim et al. / Expert Systems with Applications 38 (2011) 8488–8496

H-N. Kim et aL/ Expert Systems with Applications 38(2011 )8488-8496 8495 UTU -HITI-HVSM-O-CUM With respect to the ranking accuracy, Fig. 5 confirms that CUM ontinues outperforming the other methods on all variations of N value. For example, when N is 10, CUM obtains an RK value of 0.86 whereas uru. tl. and vsm demonstrates an rK value of 0.78, 0.71, and 0.37, respectively. The result can indicate that our 503 model provides more suitable items with a higher rank in the rec ommended set than the other methods, and consequently makes ompelling recommendations for all users. More interestingly, the simple vector model did not perform well, compared to the other methods. This result might be caused by the different degree 量量 of the density of matrices. In general, it is known that CF produces locker et al., 2004). As mentioned previously, the density of the user-item matrix R we used was 13.4% whereas that of the tag- the number of recommended items(N) item matrix Q was 3.8%. We further examined the recommendation performance for Fig 4. Comparisons of precision as the number of recommended items N increases sers who had few ratings, namely cold start users, and had lots of ratings, namely active users, in the training set. Recommender systems are generally unable to make high quality recommenda- 0.9 tions, compared to the case of active users, which is pointed out as one of the limitations. We selectively considered two groups of users who have less than 10 ratings and greater than 250 ratings. 8 And for two groups we calculated precision and ranking at top-10 i.e P@10 and rK@10)obtained by the four algorithms in order to analyze whether the differences between the two groups are sta- tistically significant or not. Fig. 6 shows the results for the cold 0.2}-是..,鲁二 start group and active group. we can see from the graphs, the results of the baseline algo- rithms demonstrated that the values of the two metrics for the cold start group were considerably low. However, our method provided quite consistent performance no matter whether they have suffi ent ratings or not. Such results were caused by the fact that the baseline methods were hard to analyze the users' propensity the number of recommended items(N) items because they did not have enough information(ratings or Fig. 5. Comparisons of ranking accuracy as the number of recommended items tags). In contrast, our collaborative user model could represent suitable preference of the users by enriching both relevant topics To experimentally evaluate the performance of top-N recom- achieved by CUM and the baseline algorithms, for the cold start mendation, we calculated PoN and RKoN of each method no mat- users, precision and ranking values of the former was found to ter whether users have sufficient ratings or not. We selectively be superior to those of the other methods. For example, CUM ob- aried the number of returned items N from 1 to 10. First, we mea- tains 10%, 15%, and 13% improvement for precision compared to ured P@N obtained by UTU. IT, VSM, and CUM UTU, ITI, and VSM, respectively. with respect to the ranking accu o Fig. 4 shows the results of precision showing how CUM outper- racy, it is clear that CUM significantly outperforms the three meth- sion values tend to decrease as the number of recommended items alleviating the problem of the cold start users and thus in p0i3 ods. This result indicates that our user model N increases. However, in the case of VSM, there is very little differ- ing the quality of item recommendations ence for different values of N. Comparing the results achieved by With respect to the active group. it can be observed that CF ap- CUM and the baseline methods, the precision value of the former proaches(i.e, UTU and ITI) provide better performance than our was found to be superior to that of the benchmark methods in method The result indicates that collaborative filtering relatively all cases. When compared to VSM, CUM is significantly more accu- works well when users have abundant rating information Compar- rate on precision. On average, on all occasions, CUM outperforms ing results for the cold start group and the active group obtained by UTU, ITI and vSM by 5%, 6. 2% and 18. 1%, respectively the Cf approaches, it is apparent that the two groups have Precision at Top 10 Ranking at Top 10 aUTU aIm DVSM O CUM 12 Iaru aIm DVSMDCUM 1.4 0.8 0.6 0.4 cold start users cold start user active users Fig. 6. Comparisons of precision and ranking accuracy for cold start users and active users

To experimentally evaluate the performance of top-N recom￾mendation, we calculated P@N and RK@N of each method no mat￾ter whether users have sufficient ratings or not. We selectively varied the number of returned items N from 1 to 10. First, we mea￾sured P@N obtained by UTU, ITI, VSM, and CUM. Fig. 4 shows the results of precision showing how CUM outper￾forms the baseline methods. The graph curves show that the preci￾sion values tend to decrease as the number of recommended items N increases. However, in the case of VSM, there is very little differ￾ence for different values of N. Comparing the results achieved by CUM and the baseline methods, the precision value of the former was found to be superior to that of the benchmark methods in all cases. When compared to VSM, CUM is significantly more accu￾rate on precision. On average, on all occasions, CUM outperforms UTU, ITI and VSM by 5%, 6.2% and 18.1%, respectively. With respect to the ranking accuracy, Fig. 5 confirms that CUM continues outperforming the other methods on all variations of N value. For example, when N is 10, CUM obtains an RK value of 0.86 whereas UTU, ITI, and VSM demonstrates an RK value of 0.78, 0.71, and 0.37, respectively. The result can indicate that our model provides more suitable items with a higher rank in the rec￾ommended set than the other methods, and consequently makes compelling recommendations for all users. More interestingly, the simple vector model did not perform well, compared to the other methods. This result might be caused by the different degree of the density of matrices. In general, it is known that CF produces good performance in situation where a rating matrix is dense (Her￾locker et al., 2004). As mentioned previously, the density of the user-item matrix R we used was 13.4% whereas that of the tag￾item matrix Q was 3.8%. We further examined the recommendation performance for users who had few ratings, namely cold start users, and had lots of ratings, namely active users, in the training set. Recommender systems are generally unable to make high quality recommenda￾tions, compared to the case of active users, which is pointed out as one of the limitations. We selectively considered two groups of users who have less than 10 ratings and greater than 250 ratings. And for two groups we calculated precision and ranking at top-10 (i.e., P@10 and RK@10) obtained by the four algorithms in order to analyze whether the differences between the two groups are sta￾tistically significant or not. Fig. 6 shows the results for the cold start group and active group. As we can see from the graphs, the results of the baseline algo￾rithms demonstrated that the values of the two metrics for the cold start group were considerably low. However, our method provided quite consistent performance no matter whether they have suffi- cient ratings or not. Such results were caused by the fact that the baseline methods were hard to analyze the users’ propensity for items because they did not have enough information (ratings or tags). In contrast, our collaborative user model could represent suitable preference of the users by enriching both relevant topics and irrelevant topics from similar neighbors. Comparing the results achieved by CUM and the baseline algorithms, for the cold start users, precision and ranking values of the former was found to be superior to those of the other methods. For example, CUM ob￾tains 10%, 15%, and 13% improvement for precision compared to UTU, ITI, and VSM, respectively. With respect to the ranking accu￾racy, it is clear that CUM significantly outperforms the three meth￾ods. This result indicates that our user model can help, indeed, in alleviating the problem of the cold start users and thus in improv￾ing the quality of item recommendations. With respect to the active group, it can be observed that CF ap￾proaches (i.e., UTU and ITI) provide better performance than our method. The result indicates that collaborative filtering relatively works well when users have abundant rating information. Compar￾ing results for the cold start group and the active group obtained by the CF approaches, it is apparent that the two groups have the number of recommended items (N) precision Fig. 4. Comparisons of precision as the number of recommended items N increases. ranking accuracy the number of recommended items (N) Fig. 5. Comparisons of ranking accuracy as the number of recommended items N increases. Precision at Top 10 Ranking at Top 10 Fig. 6. Comparisons of precision and ranking accuracy for cold start users and active users. H.-N. Kim et al. / Expert Systems with Applications 38 (2011) 8488–8496 8495

8496 H.-N. Kim et aL/ Expert Systems with Applications 38(2011)8488-8496 significant performance disparity. For example, the active users Bischoff, K Firan, C.S., Nejdl, w& Paiu, R(2008) Can all tags be used for search. In based on UTU and Iml obtain 22% and 19% improvement in terms Proceeding of the 17th ACM conference on information and knowledge of P@10, compared to the cold start users, respectively. Similar lev- Carmagnola, E, cena, E. console. L Cortassa. 0. Gena. C Coy, A et aL (2008). Tag- number of ratings is a significant factor affecting the quality of ni8(5,497-538 recommendation in CF systems. However, in practice, ev Cremonesi, P Koren, Y,& Turrin, R, 2010. Performance of recommender algorithm recommendation tasks. In Proceedings of the 4th ACM conference the contrary to the Cf approaches, there is very little difference in precision for the two groups in CUM, although it appears that Deshpande, M, Karypis, G (2004). Item-based top-n t Isers. Comparing results of RK@10 in the cold start and active Frank, E. Hall, M. & Pfahring Proceedings of the 19th conference on uncertain group achieved by CUM, interesting results were observed. The va ue of the active group was rather worse than that of the cold start Godoy, D, andi, A.(2005) User profiling in personal information agents: A for recommendations works well for survey. The knowledge Engineering Review, 20(4), 329-361 group.That is, enriched topic for tne active users, enriched topics can give negative influence on performance because they may al- Guy, L, Zwerdling. N- Ronen, L, Carmel, D.& Uziel, E(2010). So ready have satisfactory relevant and irrelevant topics for them selves. Although the performance of CUM is worse than that of nternational AM SiGir conference on research and information retrieval(pp. 194-201). produces very consistent quality of item recommendations for generation: A frequent-pattern tree approach. Data Mining and Knowledge oth the cold start users and the active users Herlocker, J. L, Konstan, J. A, Terveen, L G,& Riedl, J .(2004). Evaluating 6. Conclusions and future work Ji, A-T, Yeon, C, Kim, H. -N.& Jo, G. -S.(2007). Collaborative tagging in For the future of the Social Web, social recommender systems u x cuo. 8 Zhao ol o feren Tax(2010). Connecting users and items with to enhance the quality of recommendations In this paper, we have presented a method of building a user model incorporated with 21st ACM conference on hyperte ratings and tags. We also propose the process of topic-driven Evaluating similarity measures for emergent semantics of so enrichments to discover tags of value for users, in turn, enabling Proceedings of the 18th intemational conference on World wide Web(pp. 641 an individual model to be abundant. Empirically, the recommenda tion method based on our collaborative model outperforms the ini- McCallum, A,& Nigam, K(1998).A of event models for naive b tial model as well as the standard baseline methods. In additio the proposed approach seems to produce rather robust perfor- McLaughin, M.R.& Herlocker, J. L(200 mance in the situation where users have insufficient information model the about their preference, i. e, cold start users. promising results, it has also opened several tasks for interesting extensions. Artifacimd. S, Miva. reiviewemu3 1827-209-the-art and ging in ovic, M.(2010) Socia tags, users make frequent use of ambiguous and synonymous tags Reasonable tag-based collaborative filtering older& Huberman 2006). The recent systems, such as Faviki Proceedings of the 2nd ACM workshop on information credibility on the Web(pp. and Zigtag, support semantic tagging that allows users to assign Resnick, P. lacovo, N, Suchak, M, Bergstorm, P,& Riedl, J(1994). GroupLens: tags with well-defined concepts. By considering semantics of tags. An open architecture for collaborative filtering of netnews. In Proceedings we intend to analyze further the inside of user-generated tags in or- of 1994 ACM conference on computer supported cooperative work der to build better a user model. Second, to deal with a different sce- Sarwar, B. Karypis, G, Konstan. ].& Reidl. ](2001) Item-based collaborative ario, we intend to investigate the possible usages of our model for filtering recommendation algorithms. In Proceedings of the 10th intemational social search, which is one of emerging topics in the Social Web conference on World wide Web(pp 285-295) Schein, A L, Popescul, A, Ungar, L H, Pennock, D M. (2002). for cold-start collaborative filtering. In Proceedings of the 25th international ACM SIGIR conference on research and development in information retrieval (pp 253- References Sen, S, Vig. J ,& Riedl, J (2009). Tagommenders: Conne users to items through Adomavicius, G,& Tuzhilin, A.(2005). Toward the Next Generation of Proceedings of the 18th international conference on World wide Web(pp nsions. IEEE Transactions on Knowledge and Data Engineering. 1 Siersdorfer, S,& Sizov, S.(2009) Social recommender systems for web 2.0 734-749 oceedings of hypertext and graal, R,& Srikant, R (1994) Fast algorithms for mining association rules. In hypermedia(pp. 261-270) roceedings of 1994 international conference on very large data bases(pp. 487- Tso-Sutter, K H. L, Marinho, L B,& Thieme, L S(2008). Tag-aware recommender systems by fusion of collaborative filtering algorithms. In Proceedings of the 23rd u Yeung, C. M, Cibbins, N, Shadbolt, N, 2009. Multiple interests of users in collaborative tagging systems. In L. King &R. Baeza-Yates, (Ed wang. I. clements, M a wang. om pe vries. AS P. Reinders. M. IT.(2010, Personalization of tagging systems. In formation Processing and Management, Breese, J. S, Heckerman, D,& Wetzler nmermann, C, Bauckhage, C,& Albayrak, S(2010). I Tag, You T uncertainty in artificial intelligence(pp. 43-52) Proceedings of the 3rd ACM lm联/ /w. faviki. com g. In Proceedings of the 3rd Acm conference on recommender systems(pp ww.Zigtag. com

significant performance disparity. For example, the active users based on UTU and ITI obtain 22% and 19% improvement in terms of P@10, compared to the cold start users, respectively. Similar lev￾els of improvement can be seen on RK@10. This implies that the number of ratings is a significant factor affecting the quality of the recommendation in CF systems. However, in practice, even though users are very active, each individual has only expressed a rating on a very small portion of the total number of items. On the contrary to the CF approaches, there is very little difference in precision for the two groups in CUM, although it appears that the active users achieved slightly better results than the cold start users. Comparing results of RK@10 in the cold start and active group achieved by CUM, interesting results were observed. The va￾lue of the active group was rather worse than that of the cold start group. That is, enriched topics for recommendations works well for the cold start users. However, for the active users, enriched topics can give negative influence on performance because they may al￾ready have satisfactory relevant and irrelevant topics for them￾selves. Although the performance of CUM is worse than that of the CF methods in the active group, notably the proposed method produces very consistent quality of item recommendations for both the cold start users and the active users. 6. Conclusions and future work For the future of the Social Web, social recommender systems with tags are becoming widely used as an important technique to enhance the quality of recommendations. In this paper, we have presented a method of building a user model incorporated with ratings and tags. We also propose the process of topic-driven enrichments to discover tags of value for users, in turn, enabling an individual model to be abundant. Empirically, the recommenda￾tion method based on our collaborative model outperforms the ini￾tial model, as well as the standard baseline methods. In addition, the proposed approach seems to produce rather robust perfor￾mance in the situation where users have insufficient information about their preference, i.e., cold start users. Although the approach presented in this study has shown promising results, it has also opened several tasks for interesting future work. First, as pointed out as a common problem in free-text tags, users make frequent use of ambiguous and synonymous tags (Golder & Huberman 2006). The recent systems, such as Faviki3 and Zigtag,4 support semantic tagging that allows users to assign tags with well-defined concepts. By considering semantics of tags, we intend to analyze further the inside of user-generated tags in or￾der to build better a user model. Second, to deal with a different sce￾nario, we intend to investigate the possible usages of our model for social search, which is one of emerging topics in the Social Web. References Adomavicius, G., & Tuzhilin, A. (2005). Toward the Next Generation of Recommender Systems: A survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering, 17(6), 734–749. Agrawal, R., & Srikant, R. (1994). Fast algorithms for mining association rules. In Proceedings of 1994 international conference on very large data bases (pp. 487– 499). Au Yeung, C. M., Cibbins, N., & Shadbolt, N., 2009. Multiple interests of users in collaborative tagging systems. In I. King & R. Baeza-Yates, (Eds.), Weaving services and people on the World Wide Web (pp. 255–274). Springer-Verlag. Breese, J. S., Heckerman, D., & Kadie, C. (1998). Empirical analysis of predictive algorithms for collaborative filtering. In Proceedings of the 14th conference on uncertainty in artificial intelligence (pp. 43–52). Bischoff, K., Firan, C. S., Nejdl, W., & Paiu, R. (2008). Can all tags be used for search. In Proceeding of the 17th ACM conference on information and knowledge management (pp. 203–212). Carmagnola, F., Cena, F., Console, L., Cortassa, O., Gena, C., Goy, A., et al. (2008). Tag￾based user modeling for social multi-device adaptive guides. User Modeling and User-Adapted Interaction, 18(5), 497–538. Cremonesi, P., Koren, Y., & Turrin, R., 2010. Performance of recommender algorithms on top-n recommendation tasks. In Proceedings of the 4th ACM conference on recommender systems (pp. 39–46). De Gemmis, M., Lops, P., Semeraro, G., & Basile, P. (2008). Integrating tags in a semantic content-based recommender. In Proceedings of the 2nd ACM conference on recommender systems (pp. 163–170). Deshpande, M., & Karypis, G. (2004). Item-based top-n recommendation algorithms. ACM Transactions on Information Systems, 22(1), 143–177. Frank, E., Hall, M., & Pfahringer, B. (2003). Locally weighted naive Bayes. In Proceedings of the 19th conference on uncertainty in artificial intelligence (pp. 249– 256). Godoy, D., & Amandi, A. (2005). User profiling in personal information agents: A survey. The Knowledge Engineering Review, 20(4), 329–361. Golder, S. A., & Huberman, B. A. (2006). Usage patterns of collaborative tagging systems. Journal of Information Science, 32(2), 198–208. Guy, I., Zwerdling, N., Ronen, I., Carmel, D., & Uziel, E. (2010). Social media recommendation based on people and tags. In Proceedings of the 33rd international ACM SIGIR conference on research and development in information retrieval (pp. 194–201). Han, J., Pei, J., & Yin, Y. (2004). Mining frequent patterns without candidate generation: A frequent-pattern tree approach. Data Mining and Knowledge Discovery, 8, 53–87. Herlocker, J. L., Konstan, J. A., Terveen, L. G., & Riedl, J. (2004). Evaluating collaborative filtering recommender systems. ACM Transactions on Information Systems, 22(1), 5–53. Ji, A. -T, Yeon, C., Kim, H. -N., & Jo, G. -S. (2007). Collaborative tagging in recommender systems. In Proceedings of the 20th Australian joint conference on artificial intelligence (pp. 377–386). Li, X., Guo, L., & Zhao, Y. (2008). Tag-based social interest discovery. In Proceedings of the 17th international conference on World Wide Web (pp. 675–684). Liang, H., Xu, Y., Li, Y., Nayak, R., & Tao, X. (2010). Connecting users and items with weighted tags for personalized item recommendations. In Proceedings of the 21st ACM conference on hypertext and hypermedia (pp. 51–60). Markines, B., Cattuto, C., Menczer, F., Benz, D., Hotho, A., & Stumme, G. (2009). Evaluating similarity measures for emergent semantics of social tagging. In Proceedings of the 18th international conference on World Wide Web (pp. 641– 650). McCallum, A., & Nigam, K. (1998). A comparison of event models for naïve Bayes text classification. In Proceedings of AAAI-98 workshop on learning for text categorization (pp. 41–48). McLaughin, M. R., & Herlocker, J. L. (2004). A collaborative filtering algorithm and evaluation metric that accurately model the user experience. In Proceedings of the 27th international ACM SIGIR conference on research and development in information retrieval (pp. 329–336). Milicevic, A. K., Nanopoulos, A., & Ivanovic, M. (2010). Social tagging in recommender systems: a survey of the state-of-the-art and possible extensions. Artificial Intelligence Review, 33(3), 187–209. Nakamoto, R.Y., Nakajima, S., Miyazaki, J., Uemura, S., Kato, H., & Inagaki, Y. (2008). Reasonable tag-based collaborative filtering for social tagging systems. In Proceedings of the 2nd ACM workshop on information credibility on the Web (pp. 11–18). Resnick, P., Iacovou, N., Suchak, M., Bergstorm, P., & Riedl, J. (1994). GroupLens: An open architecture for collaborative filtering of netnews. In Proceedings of 1994 ACM conference on computer supported cooperative work (pp. 175–186). Sarwar, B., Karypis, G., Konstan, J., & Reidl, J. (2001). Item-based collaborative filtering recommendation algorithms. In Proceedings of the 10th international conference on World Wide Web (pp. 285–295). Schein, A. I., Popescul, A., Ungar, L. H., & Pennock, D. M. (2002). Methods and metrics for cold-start collaborative filtering. In Proceedings of the 25th international ACM SIGIR conference on research and development in information retrieval (pp. 253– 260). Sen, S., Vig, J., & Riedl, J. (2009). Tagommenders: Connecting users to items through tags. In Proceedings of the 18th international conference on World Wide Web (pp. 671–680). Siersdorfer, S., & Sizov, S. (2009). Social recommender systems for web 2.0 folksonomies. In Proceedings of the 20th ACM conference on hypertext and hypermedia (pp. 261–270). Tso-Sutter, K. H. L., Marinho, L. B., & Thieme, L. S. (2008). Tag-aware recommender systems by fusion of collaborative filtering algorithms. In Proceedings of the 23rd ACM symposium on applied computing (pp. 1995–1999). Wang, J., Clements, M., Wang, J., De Vries, A. P., & Reinders, M. J. T. (2010). Personalization of tagging systems. Information Processing and Management, 46(1), 58–70. Wetzker, R., Zimmermann, C., Bauckhage, C., & Albayrak, S. (2010). I Tag, You Tag: Translating tags for advanced user models. In Proceedings of the 3rd ACM international conference on web search and data mining (pp. 71–80). Zhen, Y., Li, W. J., & Yeung, D. Y. (2009). TagiCoFi: Tag informed collaborative filtering. In Proceedings of the 3rd ACM conference on recommender systems (pp. 69–76). 3 http://www.faviki.com. 4 http://www.zigtag.com. 8496 H.-N. Kim et al. / Expert Systems with Applications 38 (2011) 8488–8496