《电子商务 E-business》阅读文献：tag-aware recommender systems by fusion of collaborative filtering algorithms

Tag-aware Recommender Systems by Fusion of Collaborative Filtering Algorithms Karen H. L. Tso-Sutter, Leandro balby marinho and Lars schmidt- Thieme Information Systems and Machine Learning Lab(ISMLL University of Hildesheim Samelsonplatz 1, 31141 Hildesheim, Germany i tso, marinho, schmidt -thieme @ismll uni-hildesheimde ABSTRACT 1. INTRODUCTION Recommender Systems(RS)aim at predicting items or rat- Collaborative tagging systems, also known as folksonomies gs of items that the user are interested in. Collabora are web-based systems that allow users to upload their re- tive Filtering(CF)algorithms such as user- and item-based sources, and to label them with arbitrary words, so-called methods are the dominant techniques applied in RS algo- tags. These systems are becoming more common among web rithms. To improve recommendation quality, metadata such users now-a-day. For example popular web services such as content information of items has typically been used as Flickr del icio us, Last. fm, Gmail etc, provide pos- additional knowledge. With the increasing popularity of the sibility for users to tag or label an item of interest. In gen- collaborative tagging systems, tags could be interesting and eral, tagging is associated to the Web 2.0 and is becoming useful information to enhance rs algorithms. Unlike at the new trend enabling people to easily add metadata to tributes which are "global "descriptions of items, tags are content. Hence, These additional metadata can be used to "local"descriptions of items given by the users. To the best improve search mechanisms, better structure the data for of our knowledge, there hasn't been any prior study on tag browsing or provide personalized recommendations fitting aware RS. In this paper, we propose a generic method that he users' interests. Content information used in attribute. allows tags to be incorporated to standard CF algorithms, aware RS algorithms is typically attached to the items and is by reducing the three-dimensional correlations to three two- usually provided by domain experts. Therefore, an item al- dimensional correlations and then applying a fusion method ways has the same attributes among all users. On the other o re-associate these correlations. Additionally, we investi hand, tags are provided by various users. Thus, tags are not gate the effect of incorporating tags information to different only associated to the items but also to the users. Although CF algorithms. Empirical evaluations on three CF algo- attributes and tags are both metadata and could act as ad- rithms with real-life data set demonstrate that incorporat- ditional background knowledge to improve RS algorithms, ing tags to our proposed approach provides promising and they should be handled differently. sIg Despite the considerable amount of researches done in attribute-aware Rs algorithms, the specific problem of in Categories and Subject descriptors tegrating tags to RS algorithms is rarely explored. Most existing works on RS with tags are limited to recommend H3[Information Storage and Retrieval]: Information ing tags, i.e. assisting users for annotation purposes, while Search and Retrieval- Information Filtering using tags as supplementary source for recommending items as never been investigated General terms In this paper, we propose to integrate tags in recom- Algorithms, Performance, Experimentation mender systems by first extending the user-item matrix and then applying an algorithm that fuses two popular RS algo- rithms such that the correlations between users, items and Keywords ags can be captured simultaneously. Our contributions are Recommender Systems, Collaborative Filtering, Tags as follow: i)propose a generic method that allows tags to be incorporated to standard CF algorithms, ii) propose an "Brazilian National Council Scientific and Technological Re- dapted fusion mechanism to capture the 3-dimensional cor- search(CNPq) scholarship holder relations between users, items and tags iii)conduct empiri cal evaluations on three cF algorithms with real-life data set and investigate the effect of incorporating tags information ission to make digital or hard copies of all or part of this work for or classroom use is granted without fee provided that copies are or distributed for profit or commercial advantage and that copies notice and the full citation on the first page. To copy otherwise, te http://fickr.com republish, to post on servers or to redistribute to lists, requires prior spe 2http://del.icio.us SAC 08 March 16-20, 2008, Fortaleza, Ceara, Brazil http://www.last.fm Copyright2008ACM978-1-59593-753-7/0800035

Tag-aware Recommender Systems by Fusion of Collaborative Filtering Algorithms Karen H. L. Tso-Sutter, Leandro Balby Marinho ∗ and Lars Schmidt-Thieme Information Systems and Machine Learning Lab (ISMLL) University of Hildesheim Samelsonplatz 1, 31141 Hildesheim, Germany {tso,marinho,schmidt-thieme}@ismll.uni-hildesheim.de ABSTRACT Recommender Systems (RS) aim at predicting items or rat￾ings of items that the user are interested in. Collabora￾tive Filtering (CF) algorithms such as user- and item-based methods are the dominant techniques applied in RS algo￾rithms. To improve recommendation quality, metadata such as content information of items has typically been used as additional knowledge. With the increasing popularity of the collaborative tagging systems, tags could be interesting and useful information to enhance RS algorithms. Unlike at￾tributes which are “global” descriptions of items, tags are “local” descriptions of items given by the users. To the best of our knowledge, there hasn’t been any prior study on tag￾aware RS. In this paper, we propose a generic method that allows tags to be incorporated to standard CF algorithms, by reducing the three-dimensional correlations to three two￾dimensional correlations and then applying a fusion method to re-associate these correlations. Additionally, we investi￾gate the effect of incorporating tags information to different CF algorithms. Empirical evaluations on three CF algo￾rithms with real-life data set demonstrate that incorporat￾ing tags to our proposed approach provides promising and significant results. Categories and Subject Descriptors H.3 [Information Storage and Retrieval]: Information Search and Retrieval - Information Filtering General Terms Algorithms, Performance, Experimentation Keywords Recommender Systems, Collaborative Filtering, Tags ∗Brazilian National Council Scientific and Technological Re￾search (CNPq) scholarship holder. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SAC’08 March 16-20, 2008, Fortaleza, Ceara, Brazil ´ Copyright 2008 ACM 978-1-59593-753-7/08/0003 ...$5.00. 1. INTRODUCTION Collaborative tagging systems, also known as folksonomies are web-based systems that allow users to upload their re￾sources, and to label them with arbitrary words, so-called tags. These systems are becoming more common among web users now-a-day. For example popular web services such as Flickr 1 , del.icio.us 2 , Last.fm 3 , Gmail 4 , etc, provide pos￾sibility for users to tag or label an item of interest. In gen￾eral, tagging is associated to the Web 2.0 and is becoming the new trend enabling people to easily add metadata to content. Hence, These additional metadata can be used to improve search mechanisms, better structure the data for browsing or provide personalized recommendations fitting the users’ interests. Content information used in attribute￾aware RS algorithms is typically attached to the items and is usually provided by domain experts. Therefore, an item al￾ways has the same attributes among all users. On the other hand, tags are provided by various users. Thus, tags are not only associated to the items but also to the users. Although attributes and tags are both metadata and could act as ad￾ditional background knowledge to improve RS algorithms, they should be handled differently. Despite the considerable amount of researches done in attribute-aware RS algorithms, the specific problem of in￾tegrating tags to RS algorithms is rarely explored. Most existing works on RS with tags are limited to recommend￾ing tags, i.e. assisting users for annotation purposes, while using tags as supplementary source for recommending items has never been investigated. In this paper, we propose to integrate tags in recom￾mender systems by first extending the user-item matrix and then applying an algorithm that fuses two popular RS algo￾rithms such that the correlations between users, items and tags can be captured simultaneously. Our contributions are as follow: i) propose a generic method that allows tags to be incorporated to standard CF algorithms, ii) propose an adapted fusion mechanism to capture the 3-dimensional cor￾relations between users, items and tags iii) conduct empiri￾cal evaluations on three CF algorithms with real-life data set and investigate the effect of incorporating tags information to those algorithms. 1http://flickr.com/ 2http://del.icio.us/ 3http://www.last.fm/ 4http://mail.google.com/

2. RELATED WORK A dualistic form of user-based CF is item-based CF There have already been a reasonable amount of researches where similarities are computed between each pair of items in using attributes as background knowledge in RS 2, 3, 4 8, 7, 14, 12, 17, 16, 20. However, to the best of our knowl dge, there hasn't been any research in considering tags wit. isim(i, j) RS algorithms to predict items R'ill2IIR'iIl2 G The existing research work on using tagging information In user-based CF. to derive the recommendations for a recommendation purpo ically on the recor target user u, usually only similarities of the k most-similar mendation of tags for assisting the user in annotation re users are selected(neighborhood -Nu). When predicting a lated tasks [ 15, 22, 5, 13, 11]. The relationship between rating of a given user u for an item i, the weighted sum of users, items and tags could be treated by the literature as the other users are computed by: multidimensional R. where other dimensions of data be. sides the traditional two-dimensional user/item approach ire used. Adomavicius et al. I have proposed a reduc- ∈ Nundomr (u,v)(r(t,i)-元) tion from multidimensional to 2-dimensional representations where traditional RS algorithms can be directly applied. As- where dom r is the domain of r. For i E I we define ri pects or contexts of the additional dimensions are used to rouxii and w(u, v)is the similarity between each user u restrict or contextualize the two-dimensional data to be con- user u's neighborhood and user u him/herself. sidered, This approach is more suitable for RS that consider In the case of item-based CF, the prediction would be the time related aspect, which we do not take into account in the tag-aware RS. In our case, we use all the information average of the ratings of k most-similar items N rated by the given user u. In similar notation as user-based CF, the available in all dimensions independent of context prediction for a rating of a given user u for an item i is 3. COLLABORATIVE FILTERING Recommender systems(RS)predict ratings of items (u)=2nm(,() suggest a list of items that is unknown to the user. T ∈ Windom(i,j) ke the users, items as well as the ratings of items into a Due to its simplicity and rather promising performance count. We introduce some notations that we use throughout [10, 18), collaborative filtering has been one of the most the rest of the paper. Let prominent methods used in recommender systems 3.1 Extension with Tags U be a set of users As collaborative tagging is getting more widely used, this ite information could also be employed as background knowl- dge in RS. There has not been any previous investigation in integrating tags information to improve recommendation RERUXI be the user-item matrix where each value quality for such purpose. Hence, we propose a generic mech correspond to Ru.=r(u,i), where u∈ U and i∈I anism that allows tags to be integrated to standard CF al Two different recommendation tasks are typically consid gorithms such as user- and item-based algorithms. In the following, we will describe how the extension can be applied red:(i) predicting the ratings, i.e. how much a given user Unlike attributes which only have a two-dimensional rela- will like a particular item, and (i) predicting the items, i.e. tion item, attribute > tags hold a three-dimensional re- hich N items a user will rate, buy or visit next( topN). lation user, item, tag > We cope with this three dimen- sionalities by projecting it as three two-dimensional prob- user by using opinions from people who have alike tastes lem, user, tag and item, tag and user, item called neighborhood, while concealing the real identity of the This can be done by augmenting the standard user-item users neighborhood The prevalent method in practice is Collaborative Filter- matrix horizontally and vertically with user and item tags ing(CF)10. Its idea is basically the nearest neighbor correspondingly. User tags, are tags that user u, uses to tag items and are viewed as items in the user-item matrix. method. Given some user profiles, it predicts whether a Item tags, are tags that describe an item, i, by rs and user might be interested in a certain item based on a sec. tion of ot her users or items in the database. There are in play the role of users in the user-item matrix(See Fig. 1) Furthermore, instead of viewing each single tag as user or general two types of collaborative filtering: user-based and item, clustering methods can be applied to the tags such em-based. Most of the time, they share the same concep except they vary by how the neighborhood is formed that similar tags are grouped together. In this paper, no In user-based CF [19, recommendations are generated by clustering method has been applied to the tags. Let considering solely the ratings of users on items, by comput. ·aset ing the pairwise similarities between users, e. g, by means of vector similarity · a set of item tags be T usimating(u, v): =-(Ru,,R a new set of users after item tags extension be vertend U+T where u, vEU are two users and R and R are the a new set of items after user tags extension be Extend vectors of th I+T

2. RELATED WORK There have already been a reasonable amount of researches in using attributes as background knowledge in RS [2, 3, 4, 8, 7, 14, 12, 17, 16, 20]. However, to the best of our knowl￾edge, there hasn’t been any research in considering tags with RS algorithms to predict items. The existing research work on using tagging information for recommendation purposes lies basically on the recom￾mendation of tags for assisting the user in annotation re￾lated tasks [15, 22, 5, 13, 11]. The relationship between users, items and tags could be treated by the literature as multidimensional RS, where other dimensions of data be￾sides the traditional two-dimensional user/item approach are used. Adomavicius et al. [1] have proposed a reduc￾tion from multidimensional to 2-dimensional representations where traditional RS algorithms can be directly applied. As￾pects or contexts of the additional dimensions are used to restrict or contextualize the two-dimensional data to be con￾sidered, This approach is more suitable for RS that consider time related aspect, which we do not take into account in the tag-aware RS. In our case, we use all the information available in all dimensions independent of context. 3. COLLABORATIVE FILTERING Recommender systems (RS) predict ratings of items or suggest a list of items that is unknown to the user. They take the users, items as well as the ratings of items into ac￾count. We introduce some notations that we use throughout the rest of the paper. Let • U be a set of users, • I be a set of items, • r : U × I → R be a map of ratings, • R ∈ R U×I be the user-item matrix where each value correspond to Ru,i = r(u, i), where u ∈ U and i ∈ I. Two different recommendation tasks are typically consid￾ered: (i) predicting the ratings, i.e. how much a given user will like a particular item, and (ii) predicting the items, i.e. which N items a user will rate, buy or visit next (topN). Most recommender systems derive recommendations to a user by using opinions from people who have alike tastes, called neighborhood, while concealing the real identity of the users neighborhood. The prevalent method in practice is Collaborative Filter￾ing (CF) [10]. Its idea is basically the nearest neighbor method. Given some user profiles, it predicts whether a user might be interested in a certain item, based on a sec￾tion of other users or items in the database. There are in general two types of collaborative filtering: user-based and item-based. Most of the time, they share the same concept except they vary by how the neighborhood is formed. In user-based CF [19], recommendations are generated by considering solely the ratings of users on items, by comput￾ing the pairwise similarities between users, e.g., by means of vector similarity: usimratings(u, v) := hRu,., Rv,.i ||Ru,.||2||Rv,.||2 (1) where u, v ∈ U are two users and Ru,. and Rv,. are the vectors of their ratings. A dualistic form of user-based CF is item-based CF [9], where similarities are computed between each pair of items i, j ∈ I. isimratings(i, j) := hR 0 .,i, R0 .,j i ||R0 .,i||2||R0 .,j ||2 (2) In user-based CF, to derive the recommendations for a target user u, usually only similarities of the k most-similar users are selected (neighborhood – Nu). When predicting a rating of a given user u for an item i, the weighted sum of the other users are computed by: rˆ(u, i) := ¯ru + P v∈Nu∩domri w(u, v)(r(v, i) − r¯v) P v∈Nu∩domri w(u, v) (3) where dom r is the domain of r. For i ∈ I we define ri := r|U×{i} and w(u, v) is the similarity between each user v in user u’s neighborhood and user u him/herself. In the case of item-based CF, the prediction would be the average of the ratings of k most-similar items Ni rated by the given user u. In similar notation as user-based CF, the prediction for a rating of a given user u for an item i is: rˆ(u, i) := P j∈Ni∩domru w(i, j)(r(i, j)) P j∈Ni∩domru w(i, j) (4) Due to its simplicity and rather promising performances [10, 18], collaborative filtering has been one of the most prominent methods used in recommender systems. 3.1 Extension with Tags As collaborative tagging is getting more widely used, this information could also be employed as background knowl￾edge in RS. There has not been any previous investigation in integrating tags information to improve recommendation quality for such purpose. Hence, we propose a generic mech￾anism that allows tags to be integrated to standard CF al￾gorithms such as user- and item-based algorithms. In the following, we will describe how the extension can be applied. Unlike attributes which only have a two-dimensional rela￾tion , tags hold a three-dimensional re￾lation . We cope with this three dimen￾sionalities by projecting it as three two-dimensional prob￾lem, and and . This can be done by augmenting the standard user-item matrix horizontally and vertically with user and item tags correspondingly. User tags, are tags that user u, uses to tag items and are viewed as items in the user-item matrix. Item tags, are tags that describe an item, i, by users and play the role of users in the user-item matrix (See Fig. 1). Furthermore, instead of viewing each single tag as user or item, clustering methods can be applied to the tags such that similar tags are grouped together. In this paper, no clustering method has been applied to the tags. Let: • a set of user tags be Tu, • a set of item tags be Ti, • a new set of users after item tags extension be Uextend = U + Ti, • a new set of items after user tags extension be Iextend = I + Tu

similar users. In this paper, we only consider the first type of combination as our initial experiments have shown that the second type did not provide better recommendations in he predicting item problem. The fusion of the user- and item-based predictions was done by computing the sum of he two conditional probabilities that are based on user-and item-based similarities, which are computed using standard user-and item-based CF. A parameter, A, is introduced to adjust the significance of the two predictions P(ru,ilw(u, u))A+ P(Tuilw(i, a)(1-A) R In this paper, we refer the combination of user- and item- 3.2.1 Fusion for Predicting Item Problem Most systems that use collaborative tagging do not con- ain rating information, i.e. only the occurrence, Ou.iE 0, 1, whether item i E I occurred with user uEU(e.g. u has bought/viewed item i). For example, Last. fm is a popu- lar internet radio and music community website which allows user to tag the music. However, it does not support user Figure 1: Extend user-item matrix by including user to provide explicit ratings of the music. A predicting item tags as items and item tags as users. problem would be more suitable for tag-aware RS. Hence, we focus on the predicting item problem in this paper. Yet fu E (0, 1 specify whether a user u E U has used the fusion method by Wang et 1 does not consider tags. Also, their algorithm is only suitable for predicting rating problem and not the predicting item problem. Thus ·O:4∈{0,1}, specify whether an item i∈ i is de we propose a fusion algorithm that tackles the predicting scribed by the tag ti, where ti E T item problem and also takes tags into account RT, be represented in a UxITu user-tag matrix, where Again, fusion for the predicting item problem is done by each value of the matrix corresponds to Ou, t combining the predictions of user and item-based; however, this time, the predictions are computed differently. For the Rr, be represented in a Ti x I tag-item matrix, where predicting item problem in user-based CF, recommendations each value of the matrix corresponds to Oi, tt are a list of items that is ranked by decreasing frequency of occurrence in the ratings of his/her neighbors To apply user-and item-based CF after the extension with ags, both CF algorithms have to be recomputed with the (Oa,=1) {u∈N|On=1H newly extended user-item matrix. For user-based CF, the (5) new user-item matrix, Ruertend: = R+ RTu, is represented n a ux lertend matrix. In the case of item-based CF. the For item-based CF, the topN recommendation suggested new user-item matrix, Riertend: =R+Rr, is represented by [9] is to compute a list of items that is ranked by decreas- in a Extend x I matrix. ing sum of the similarities of neighboring items, Ni, which have been rated by user u 3.2 Fusing User-based and item-based p(Ont=1)=∑t(,j Tag information has a slightly different nature than at j∈N4nOu,j tribute information. In general, attributes are only attached o items and attributes of an item appear the same to all As the values of the prediction lists computed by user- users"globally". Tags of an item, on the other hand, are de- item-based have different units(user-based being the riptions of the item by one or more than one users. Thus frequency of items and item-based the similarity of items) To address the different meanings of the values, we normal- depended on the user's preference. They are "local"de- ized the prediction lists to unity. Note that other ways of criptions of an item that might change from users to users This suggests that a RS algorithm that is able to capture we have chosen to normalize to unity both user's and items aspect of tags would eventually be a The topN combined prediction list is thus suitable choice We have selected an existing algorithm developed by Wang 1) et al. [21], which fuses the predictions of user- and item (Ou,=1):=A based CF. In general, their idea is to correspond Cf to the estimation of conditional probability problem. Their paper described two types of combination: i) predicting the rating by fusing user- and item-based predictions; ii)in addition +(1-入) to i), it also used the similar items ratings generated by p(On,=1)

Figure 1: Extend user-item matrix by including user tags as items and item tags as users. • Ou,tu ∈ {0, 1} specify whether a user u ∈ U has used the tag tu, where tu ∈ Tu, • Oi,ti ∈ {0, 1}, specify whether an item i ∈ I is de￾scribed by the tag ti, where ti ∈ Ti, • RTu be represented in a U×|Tu| user-tag matrix, where each value of the matrix corresponds to Ou,tu , • RTi be represented in a |Ti|×I tag-item matrix, where each value of the matrix corresponds to Oi,ti . To apply user- and item-based CF after the extension with tags, both CF algorithms have to be recomputed with the newly extended user-item matrix. For user-based CF, the new user-item matrix , Ruextend := R + RTu , is represented in a U × Iextend matrix. In the case of item-based CF, the new user-item matrix, Riextend := R + RTi , is represented in a Uextend × I matrix. 3.2 Fusing User-based and Item-based Tag information has a slightly different nature than at￾tribute information. In general, attributes are only attached to items and attributes of an item appear the same to all users “globally”. Tags of an item, on the other hand, are de￾scriptions of the item by one or more than one users. Thus, tags are not only attached to the item itself but also are depended on the user’s preference. They are “local” de￾scriptions of an item that might change from users to users. This suggests that a RS algorithm that is able to capture both user’s and item’s aspect of tags would eventually be a suitable choice. We have selected an existing algorithm developed by Wang et al. [21], which fuses the predictions of user- and item￾based CF. In general, their idea is to correspond CF to the estimation of conditional probability problem. Their paper described two types of combination: i) predicting the rating by fusing user- and item-based predictions; ii) in addition to i), it also used the similar items ratings generated by similar users. In this paper, we only consider the first type of combination as our initial experiments have shown that the second type did not provide better recommendations in the predicting item problem. The fusion of the user- and item-based predictions was done by computing the sum of the two conditional probabilities that are based on user- and item-based similarities, which are computed using standard user- and item-based CF. A parameter, λ, is introduced to adjust the significance of the two predictions. P(ru,i|w(u, v))λ + P(ru,i|w(i, z))(1 − λ). In this paper, we refer the combination of user- and item￾based CF as fusion. 3.2.1 Fusion for Predicting Item Problem Most systems that use collaborative tagging do not con￾tain rating information, i.e. only the occurrence, Ou,i ∈ {0 , 1 }, whether item i ∈ I occurred with user u ∈ U (e.g. u has bought/viewed item i). For example, Last.fm is a popu￾lar internet radio and music community website which allows user to tag the music. However, it does not support users to provide explicit ratings of the music. A predicting item problem would be more suitable for tag-aware RS. Hence, we focus on the predicting item problem in this paper. Yet, the fusion method by Wang et al. [21] does not consider tags. Also, their algorithm is only suitable for predicting rating problem and not the predicting item problem. Thus, we propose a fusion algorithm that tackles the predicting item problem and also takes tags into account. Again, fusion for the predicting item problem is done by combining the predictions of user and item-based; however, this time, the predictions are computed differently. For the predicting item problem in user-based CF, recommendations are a list of items that is ranked by decreasing frequency of occurrence in the ratings of his/her neighbors. p ucf(Ou,i = 1) := |{v ∈ Nu | Ov,i = 1}| |Nu| (5) For item-based CF, the topN recommendation suggested by [9] is to compute a list of items that is ranked by decreas￾ing sum of the similarities of neighboring items, Ni, which have been rated by user u. p icf(Ou,i = 1) := ❳ j∈Ni∩Ou,j=1 w(i, j) (6) As the values of the prediction lists computed by user￾and item-based have different units (user-based being the frequency of items and item-based the similarity of items). To address the different meanings of the values, we normal￾ized the prediction lists to unity. Note that other ways of normalization techniques can be explored but for simplicity we have chosen to normalize to unity. The topN combined prediction list is thus: p iucf(Ou,i = 1) := λ p ucf ❳ (Ou,i = 1) i p ucf (Ou,i = 1) +(1 − λ) p icf ❳ (Ou,i = 1) i p icf (Ou,i = 1) (7)

Predctxe pefromance dingus r Fusion with Tag As user- and item-based CF are the basic components of the fusion method, when tag information is available, fusion can simply be extended with tags by applying the tag extension method described in Section 3.1 We deem that this tag-fusion algorithm is a suitable tag aware RS algorithm because i) user tags and item tags pro- vide extra indications of the user's and item's preferences our adapted fusion approach then brings about both user d item aspects of the tags concurrently. In fact, our em- Figure 2: Results of comparing baseline models and pirical analysis has shown that our tag-fusion RS approach the unification model using the optimized lambda provides promising results. and neighborhood. 4. EXPERIMENTS items to be predicted is 10 4.1 Data Set Methods in Comparison We have implemented two baseline models: user-based CF The data set we have used in our experiments is the [19] and item-based CF [9 as well as the fusion of user-and Last. fm data. Last. fm is an internet radio and music com item-based cF described in Section 3. 2.1. In addition. to munity website and is one of the world's largest social music evaluate the impact of the presence of tags in CF algorithms platforms. It collects the profile of each user's musical pref we have compared the evaluations with the incorporation of rence by keeping track of the songs that the user listens tags and without tags to each algorithm. Hence, each al 0. The reason we have chosen this data set is because it is gorithm is first evaluated with the standard approach and rich in users preference information-songs the users then evaluated with the extension of tags using the method to, as well as tags information- user-end tagging of artists, described in Section 3. 1. Note that all algorithms are im- albums, and tracks of the music. The data was gathered plemented for the predicting item problem and all Cf algo- rithms are computed by the vector similarity measures. re artist names, which are already normalized by the sy em. Many recommendation algorithms suffer from sparse 4.3 Experiment Results data or the "long tail"of items which were used by only few Figure 2 summarizes the results, including a comparison sers. Hence, to eliminate some of the noise and improve with the baseline models: standard user- and item-based CE the chances of good results for all algorithms we will adopt as well as the fusion model, each with and without the pres- the pruning procedure described in [11 and restrict the eval- ence of tags. It can be seen that the fusion method,both uation to the dense"part of the tripartite graph. In this with and without tags, significantly out perform the stan- case we considered only the users, resources and tags that dard CF models. Furthermore. the results show that after appear at least 10 times in the tag assignments, i.e. (user, the introduction of tags to the fusion method, there is a sig esource,tag)triples. This gives 1853 items(artists), 2917 nificant increase in the performance. It is interesting to see users and 2045 tags that incorporating tags to the baseline models does not im- 4.2 Evaluation Protocol and setting prove the recommendation quality at all, in contrast to the promising results of including tags in the fusion method. As We have used a leave-one-out protocol to evaluate the ob- mentioned in previous section, tags hold the 3-dimensional tained recommendations. Hence, obtaining the test set by relationship between users, items and tags. The use of tags randomly select one listened music from every user. The rest has shown to increase the interconnectivity amongst users f the data is used as the training set, where the model trained on this data set. This protocol has been called All- the characteristic of tags correctly. Hence, attaching tags to ButI in[6. We have performed 10-fold cross validation. In standard CF algorithms, such as user-based CF, does not addition, we have further split the training data to validation improve the performance at all, the tags are then only seen data to optimize the parameters X and k, the neighborhoo as noise. This reflects that by simply extending the standard size. We have varied the lambda from 0 to 1 by an interval CF algorithms with tags, it fails to denote the 3-dimensional of 0. I and the neighborhood from 10-150 by an interval of correlations between user, item and tag, whereas the pro- 10. Using the validation data, we have found the best x to posed fusion method has shown to be able to capture this be 0.4 and k to be 20. We have then retrained the model relationship using the training data with the optimized lambda. Our paper focuses on the item prediction problem, which is to predict a fixed number of topN recommendations an 5. CONCLUSION ot the ratings. Suitable evaluation metrics are Precision We have conducted a novel study in the utilization of tag Recall and F1. Similar to Sarwar et a our evaluations ementary source to predict item recommendations. ider any item in the recommendation set that matches e have presented a generic method to include tags to item in the testing set as ahit". The number of topN st CF algorithms such as user- and item-based CF. In

arg Nmax i p iucf(Ou,i = 1) (8) Fusion with Tags As user- and item-based CF are the basic components of the fusion method, when tag information is available, fusion can simply be extended with tags by applying the tag extension method described in Section 3.1. We deem that this tag-fusion algorithm is a suitable tag￾aware RS algorithm because i) user tags and item tags pro￾vide extra indications of the user’s and item’s preferences, ii) our adapted fusion approach then brings about both user and item aspects of the tags concurrently. In fact, our em￾pirical analysis has shown that our tag-fusion RS approach provides promising results. 4. EXPERIMENTS 4.1 Data Set The data set we have used in our experiments is the Last.fm data. Last.fm is an internet radio and music com￾munity website and is one of the world’s largest social music platforms. It collects the profile of each user’s musical pref￾erence by keeping track of the songs that the user listens to. The reason we have chosen this data set is because it is rich in users preference information – songs the users listen to, as well as tags information – user-end tagging of artists, albums, and tracks of the music. The data was gathered in July 2006 by crawling the Last.fm site. Here the items are artist names, which are already normalized by the sys￾tem. Many recommendation algorithms suffer from sparse data or the “long tail” of items which were used by only few users. Hence, to eliminate some of the noise and improve the chances of good results for all algorithms we will adopt the pruning procedure described in [11] and restrict the eval￾uation to the “dense” part of the tripartite graph. In this case we considered only the users, resources and tags that appear at least 10 times in the tag assignments, i.e. (user, resource, tag) triples. This gives 1853 items (artists), 2917 users and 2045 tags. 4.2 Evaluation Protocol and Setting We have used a leave-one-out protocol to evaluate the ob￾tained recommendations. Hence, obtaining the test set by randomly select one listened music from every user. The rest of the data is used as the training set, where the model is trained on this data set. This protocol has been called All￾But1 in [6]. We have performed 10-fold cross validation. In addition, we have further split the training data to validation data to optimize the parameters λ and k, the neighborhood size. We have varied the lambda from 0 to 1 by an interval of 0.1 and the neighborhood from 10-150 by an interval of 10. Using the validation data, we have found the best λ to be 0.4 and k to be 20. We have then retrained the model using the training data with the optimized lambda. Our paper focuses on the item prediction problem, which is to predict a fixed number of topN recommendations and not the ratings. Suitable evaluation metrics are Precision, Recall and F1. Similar to Sarwar et al. [19], our evaluations consider any item in the recommendation set that matches any item in the testing set as a “hit”. The number of topN Figure 2: Results of comparing baseline models and the unification model using the optimized lambda and neighborhood. items to be predicted is 10. Methods in Comparison We have implemented two baseline models: user-based CF [19] and item-based CF [9] as well as the fusion of user- and item-based CF described in Section 3.2.1. In addition, to evaluate the impact of the presence of tags in CF algorithms, we have compared the evaluations with the incorporation of tags and without tags to each algorithm. Hence, each al￾gorithm is first evaluated with the standard approach and then evaluated with the extension of tags using the method described in Section 3.1. Note that all algorithms are im￾plemented for the predicting item problem and all CF algo￾rithms are computed by the vector similarity measures. 4.3 Experiment Results Figure 2 summarizes the results, including a comparison with the baseline models: standard user- and item-based CF as well as the fusion model, each with and without the pres￾ence of tags. It can be seen that the fusion method, both with and without tags, significantly outperform the stan￾dard CF models. Furthermore, the results show that after the introduction of tags to the fusion method, there is a sig￾nificant increase in the performance. It is interesting to see that incorporating tags to the baseline models does not im￾prove the recommendation quality at all, in contrast to the promising results of including tags in the fusion method. As mentioned in previous section, tags hold the 3-dimensional relationship between users, items and tags. The use of tags has shown to increase the interconnectivity amongst users and items. Applying user/item tags alone does not exploit the characteristic of tags correctly. Hence, attaching tags to standard CF algorithms, such as user-based CF, does not improve the performance at all, the tags are then only seen as noise. This reflects that by simply extending the standard CF algorithms with tags, it fails to denote the 3-dimensional correlations between user, item and tag, whereas the pro￾posed fusion method has shown to be able to capture this relationship. 5. CONCLUSIONS We have conducted a novel study in the utilization of tags as supplementary source to predict item recommendations. Here, we have presented a generic method to include tags to standard CF algorithms such as user- and item-based CF. In

addition, we have found an approach that deals with the 3- L. Schmidt-Thieme and G. Stumme. Tag dimensional correlation between the users, items and tags by recommendations in folksonomies. In Proceedings of first applying our tag extension mechanism and then an fu- the 11th European Conference on Principles and sion method which we have adapted from a predicting rating Practice of Knowledge Discovery in Databases roblem to predicting item problem. Our empirical analysis (PKDD), Warsaw, Poland(to appear), 2007. has shown that the proposed adapted fusion method outper- [12 Q. Lin and B M. Kim. An approach for combining forms standard baseline models, especially with the incorpo- content-based and collaborative filters. In proceeding ration of tags. Moreover, our findings have suggested that of the Sixth international workshop on Information our adapted fusion method has successfully captured the retrieval with Asian languages(ACL-2003), pages relationships between users, items and tags. Although our investigation has provided promising results, we believe that [13] L. B Marinho and L.Schmidt-Thieme.Collaborative our contribution is an initial step in the study of tag-aware tag recommendations. In Proceedings of 31st Annual s, additional research in this field is still to be explored Conference of the Gesellschaft fr Klassifikation (GfKI), Freiburg(to " Springer, 2007 6. REFERENCES [14 P. Melville, R. Mooney, and R. Nagarajar [1 G. Adomavicius, R Sankaranarayanan, S Sen, and Content-boosted collaborative filtering. In Proceedings A. Tuzhilin. Incorporating contextual information in of Eighteenth National Conference on Artificia recommender systems using a multidimensional ntelligence(AAA1-2002), pages 187-192, 2001 approach. ACM Trans. Inf. Syst., 23(1): 103-145, 2005. [15] G Mishne. Autotag: a collaborative approach to 2 M. Balabanovic and Y Shoham Fab: Content-based automated tag assignment for weblog posts. In www collaborative recommendation. Commun. ACM 06: Proceedings of the 15th international conference 40:66-72,1997 on World Wide Web, pages 953-954, New York, NY 3 J. Basilico and T. Hofmann. Unifying collaborative USA, 2006. ACM Press. and content-based filtering. In Proceedings of the [16 M.J. Pazzani. A framework for collaborative twenty-first international conference on Machine content-based and demographic filtering. Artif. Intell earning, pages 65-72, New York, NY, USA, 2004. Re,13(5-6):393-408,199 ACM Press 17 A Popescul, L. Ungar, D. Pennock, and S Lawrence. 4 C Basu, H. Hirsh, and W. Cohen. Recommendation Probabilistic models for unified collaborative and as classification: using social and content-based content-based recommendation in sparse-data information in recommendation. In AAAI 98/IAAl environments. In Proceedings of the Seventeent 98:Proceedings of the fifteenth national/tenth onference on Uncertainty in Artificial Intelligence, conference on Artificial intelligence/Innovative pages437-444,2001. applications of artificial intelligence, pages 714-720 [18 P. Resnick, N. Iacovou, M. Suchak, P. Bergstorm, and Menlo Park, CA, USA, 1998. American Association J. Riedl. GroupLens: an open architecture for for Artificial Intelligence collaborative filtering of netnews. In Proceedings of the 5 D. Benz, K. Tso, and L. Schmidt-Thieme Automatic ACM1994 Conference on Computer Supported bookmark classification: A collaborative approach. In Cooperative Work, pages 175-186, Chapel Hill, North Proceedings of the Second Workshop on Innovations in Carolina. 1994. ACM Web Infrastructure(IWI 2006), Edinburgh, Scotland, [19 B M. Sarwar, G. Karypis, J. A. Konstan, and 2006 J. Riedl. Analysis of recommendation algorithms for [6J. S. Breese, D. Heckerman, and C. Kadie. Empirical e-commerce. In Proceedings of the Second ACM analysis of predictive algorithms for collaborative Conference on Electromic Commerce(EC 00), pages filtering. In Proceedings of the 14th Conference on 285-295,2000 Uncertainty in Artificial Intelligence(UAI-98). G. F. [20] K. Tso and L. Schmidt-Thieme. Evaluation of Cooper, and S. Moral, Eds. Morgan-Kaufmann, San attribute-aware recommender system algorithms on francisco, Calif, pages 43-52, 1998. data with varying characteristics. In Proceedings of the 7 R. Burke. Hybrid recommender systems: Survey and 10th Pacific-Asia Conference on Knowledge Discovery xperiments. User Modeling and User-Adapted and Data Mining(PAKDD 2006) pages 831-840 nteraction,12:331-370,2002 Springer, 2006 [8 M. Claypool, A. Gokhale, T. Miranda, P. Murnikov 21 J. Wang, A P de Vries, and M. J. T Reinders D. Netes, and M. Sartin. Combining content-based Unifying user-based and item-based collaborative and collaborative filters in an online newspaper. In filtering approaches by similarity fusion In SIGIR 06: Proceedings of ACM SIGIR Workshop on Proceedings of the 29th annual international ACM Recommender Systems, August 1999. SIGIR conference on Research and development in ⑨ de and G. Karypis. Item- information retrieval, pages 501-508, New York, NY Information Systems. Springer-Verlag, 22/1, 200 USA, 2006. ACM Pres 10 D. Goldberg, D. Nichols, B. M. Oki, and D. Terry Using collaborative filtering to weave an information Proceedings of the Collaborative Web Tagging tapestry. Communications of the ACM, 35(12): 61-70 Workshop at the www 2006, Edinburgh, Scotland 2006 [11 R Jaeschke, L Marinho, A Hotho

addition, we have found an approach that deals with the 3- dimensional correlation between the users, items and tags by first applying our tag extension mechanism and then an fu￾sion method which we have adapted from a predicting rating problem to predicting item problem. Our empirical analysis has shown that the proposed adapted fusion method outper￾forms standard baseline models, especially with the incorpo￾ration of tags. Moreover, our findings have suggested that our adapted fusion method has successfully captured the relationships between users, items and tags. Although our investigation has provided promising results, we believe that our contribution is an initial step in the study of tag-aware RS, additional research in this field is still to be explored. 6. REFERENCES [1] G. Adomavicius, R. Sankaranarayanan, S. Sen, and A. Tuzhilin. Incorporating contextual information in recommender systems using a multidimensional approach. ACM Trans. Inf. Syst., 23(1):103–145, 2005. [2] M. Balabanovic and Y. Shoham. Fab: Content-based, collaborative recommendation. Commun. ACM, 40:66–72, 1997. [3] J. Basilico and T. Hofmann. Unifying collaborative and content-based filtering. In Proceedings of the twenty-first international conference on Machine learning, pages 65–72, New York, NY, USA, 2004. ACM Press. [4] C. Basu, H. Hirsh, and W. Cohen. Recommendation as classification: using social and content-based information in recommendation. In AAAI ’98/IAAI ’98: Proceedings of the fifteenth national/tenth conference on Artificial intelligence/Innovative applications of artificial intelligence, pages 714–720, Menlo Park, CA, USA, 1998. American Association for Artificial Intelligence. [5] D. Benz, K. Tso, and L. Schmidt-Thieme. Automatic bookmark classification: A collaborative approach. In Proceedings of the Second Workshop on Innovations in Web Infrastructure (IWI 2006), Edinburgh, Scotland, 2006. [6] J. S. Breese, D. Heckerman, and C. Kadie. Empirical analysis of predictive algorithms for collaborative filtering. In Proceedings of the 14th Conference on Uncertainty in Artificial Intelligence (UAI-98). G. F. Cooper, and S. Moral, Eds. Morgan-Kaufmann, San Francisco, Calif, pages 43–52, 1998. [7] R. Burke. Hybrid recommender systems: Survey and experiments. User Modeling and User-Adapted Interaction, 12:331–370, 2002. [8] M. Claypool, A. Gokhale, T. Miranda, P. Murnikov, D. Netes, and M. Sartin. Combining content-based and collaborative filters in an online newspaper. In Proceedings of ACM SIGIR Workshop on Recommender Systems, August 1999. [9] M. Deshpande and G. Karypis. Item-based top-n recommendation algorithms. ACM Transactions on Information Systems. Springer-Verlag, 22/1, 2004. [10] D. Goldberg, D. Nichols, B. M. Oki, and D. Terry. Using collaborative filtering to weave an information tapestry. Communications of the ACM, 35(12):61–70, 1992. [11] R. Jaeschke, L. Marinho, A. Hotho, L. Schmidt-Thieme, and G. Stumme. Tag recommendations in folksonomies. In Proceedings of the 11th European Conference on Principles and Practice of Knowledge Discovery in Databases (PKDD), Warsaw, Poland (to appear), 2007. [12] Q. Lin and B. M. Kim. An approach for combining content-based and collaborative filters. In Proceedings of the Sixth international workshop on Information retrieval with Asian languages (ACL-2003), pages 17–24, 2003. [13] L. B. Marinho and L. Schmidt-Thieme. Collaborative tag recommendations. In Proceedings of 31st Annual Conference of the Gesellschaft fr Klassifikation (GfKl), Freiburg (to appear). Springer, 2007. [14] P. Melville, R. Mooney, and R. Nagarajan. Content-boosted collaborative filtering. In Proceedings of Eighteenth National Conference on Artificial Intelligence (AAAI-2002), pages 187–192, 2001. [15] G. Mishne. Autotag: a collaborative approach to automated tag assignment for weblog posts. In WWW ’06: Proceedings of the 15th international conference on World Wide Web, pages 953–954, New York, NY, USA, 2006. ACM Press. [16] M. J. Pazzani. A framework for collaborative, content-based and demographic filtering. Artif. Intell. Rev., 13(5-6):393–408, 1999. [17] A. Popescul, L. Ungar, D. Pennock, and S. Lawrence. Probabilistic models for unified collaborative and content-based recommendation in sparse-data environments. In Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence, pages 437–444, 2001. [18] P. Resnick, N. Iacovou, M. Suchak, P. Bergstorm, and J. Riedl. GroupLens: an open architecture for collaborative filtering of netnews. In Proceedings of the ACM1994 Conference on Computer Supported Cooperative Work, pages 175–186, Chapel Hill, North Carolina, 1994. ACM. [19] B. M. Sarwar, G. Karypis, J. A. Konstan, and J. Riedl. Analysis of recommendation algorithms for e-commerce. In Proceedings of the Second ACM Conference on Electronic Commerce (EC’00), pages 285–295, 2000. [20] K. Tso and L. Schmidt-Thieme. Evaluation of attribute-aware recommender system algorithms on data with varying characteristics. In Proceedings of the 10th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD 2006), pages 831–840. Springer, 2006. [21] J. Wang, A. P. de Vries, and M. J. T. Reinders. Unifying user-based and item-based collaborative filtering approaches by similarity fusion. In SIGIR ’06: Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pages 501–508, New York, NY, USA, 2006. ACM Press. [22] Z. Xu, Y. Fu, J. Mao, and D. Su. Towards the semantic web: Collaborative tag suggestions. In Proceedings of the Collaborative Web Tagging Workshop at the WWW 2006, Edinburgh, Scotland, 2006