#BM、E如E sented in [14. After the user enters a tag for a new resource, algorithm recommends tags based on co-occurence of tags for resources which the user or others used together in the past. After each tag the user assigns or selects, the set 0.221 is narrowed down to make the tags more specific. In 27 0.091 0.241 Shepitsen et al. propose a resource recommendation system 0.105 0.256 based on hierarchical clustering of the tag space. The rec- ommended resources are identified using user profiles and #BM w/o Extended w/ Extended tag clusters to personalize the recommendation results. US- 0.105 ing LDa topic models to recommend resources rather than 234 0.039 0.15 tags is subject for future work An approach to collective tag recommendation based on association rule mined from the resource tag matrix has been introduced in 18. As discussed in Section 3.2. this a proach recommends rather general tags with low TFID and achieves smaller recall and precision than the approach Table 9: Mean Average Precision(MAP)for tag based on LDA introduced in this paper. When content of search with and without extention of recommended resources is available, tag recommendation can also be ap- tags proached as a classification problem, predicting tags from content. A recent approach in this direction is presented in 29. They cluster the document-term-tag matrix after an 4. RELATED WORK approximate dimensionality reduction, and obtain a ranked Tag recommendation has received considerable interest in membership of tags to clusters. Tags for new resources are recent years. Most work has focused on personalized tag recommended by classifying the resources into clusters, and ecommendation, suggesting tags to the user bookmarking ranking the cluster tags accordingly. a new resource: This is often done using collaborative filter- Tags have been proven to be very useful for search: in ing, taking into account similarities between users, resources case of image search where content based features are very and tags. 25 introduces an approach to recommend tags difficult to extract [12, in case of enterprise search where not enough link information is available [13, or in case of web for weblogs, based on similar weblogs tagged by the same search to optimize results [3]. A large scale evaluation of user. Chirita et al. [11] realize this idea for the personal Delicious regarding search is presented in [17]. They found desktop, recommending tags for web resources by retrieving and ranking tags from similar documents on the desktop that 50% of the pages annotated by a particular tag contain 31 aims at recommending a few descriptive tags to users the tag within the page's content. Bischoff et al. 9 provide y rewarding co-occuring tags that have been assigned by an indepth analysis of a number of tagging systems with the same user, penalizing co-occuring tags that have been respect toto their usefulness for search. They observe that assigned by different users, and boosting tags with high de descriptive tags such as topic or type tags are much more criptiveness(TFIDF). As pointed out in Section 2.2, penal- frequent than personal tags such as to read", especially in izing co-occuring tags assigned by different users in an effort the mid and low tag frequency range, and that these tags are to recommend personalized tags is in contrast to using tag indeed used in search. Berendt and Hanser [6 argue that association rules to recommend general tags to improve re- tags can be considered content and not just metadata which call for search. Sigurbjornsson and van Zwol [28 also look at makes then valuable in a content based document retrieval tags to recommend tags based on a user de- well fined set of tags. The co-occuring tags are then ranked and Recently a number of papers deal with imroving search in promoted based on e. g. descriptiveness. Jaeschke et al. 20 tagging systems. Krestel and Chen [21 propose a method compare two variants of collaborative filtering and Folkrank to measure the quality of tags with respect to the annotated a graph based algorithm for personalized tag recommenda- resource to identify high quality tags that describe a re- tion. For collaborative filtering, once the similarity between source better than others. Hotho et al. [19 propose exploit- sers on tags, and once the similarity between users on re- of ources is used for recommendation. Folkrank uses randor and ranking within tagging systems. Folk Rank"is using walk techniques on the user-resource-tag(URT) graph based graph model to represent the folksonomy and can be used on the idea that popular users, resources, and tags can re inforce each other. These algorithms take co-occurrence of present a tag clustering algorithm to improve search. The tags into account only indirectly, via the URT graph. Syme- setting is similar to ours: Related tags are identified that can nidis et al. 30 employ dimensionality reduction to per e used for extending existing resource annotations, query nalized tag recommendation. Whereas[ 20] operate on the expansion or result clustering. The clustering is based on RT graph directly, 130] use generalized techniques of SVD simple co-occurence counts. Unfortunately, the paper does (Singular Value Decomposition) for n-dimensional tensors not contain a sound evaluation of the results. Schenkel et The 3 dimensional tensor corresponding to the UrT graph al. 26 propose to improve search in tagging systems by ex- is unfolded into 3 matrices, which are reduced by means of panding a user query with semantically similar tags and rank SVD individually, and combined again to arrive at a more the result additionally based on a social component, which means that tagging information of friends of a user in the recommendation then suggests tags to users, if their weight network is taken into account when a user submits a query is above some threshold. An interactive approach is pre-#BM MAP for top 10 w/o Extended w/ Extended 1 0.037 0.137 2 0.058 0.196 3 0.075 0.221 4 0.091 0.241 5 0.105 0.256 #BM MAP for top 20 w/o Extended w/ Extended 1 0.025 0.105 2 0.039 0.150 3 0.051 0.170 4 0.062 0.186 5 0.072 0.198 Table 9: Mean Average Precision (MAP) for tag search with and without extention of recommended tags 4. RELATED WORK Tag recommendation has received considerable interest in recent years. Most work has focused on personalized tag recommendation, suggesting tags to the user bookmarking a new resource: This is often done using collaborative filter￾ing, taking into account similarities between users, resources, and tags. [25] introduces an approach to recommend tags for weblogs, based on similar weblogs tagged by the same user. Chirita et al. [11] realize this idea for the personal desktop, recommending tags for web resources by retrieving and ranking tags from similar documents on the desktop. [31] aims at recommending a few descriptive tags to users by rewarding co-occuring tags that have been assigned by the same user, penalizing co-occuring tags that have been assigned by different users, and boosting tags with high de￾scriptiveness (TFIDF). As pointed out in Section 2.2, penal￾izing co-occuring tags assigned by different users in an effort to recommend personalized tags is in contrast to using tag association rules to recommend general tags to improve re￾call for search. Sigurbj¨ornsson and van Zwol [28] also look at co-occurence of tags to recommend tags based on a user de- fined set of tags. The co-occuring tags are then ranked and promoted based on e.g. descriptiveness. Jaeschke et al. [20] compare two variants of collaborative filtering and Folkrank, a graph based algorithm for personalized tag recommenda￾tion. For collaborative filtering, once the similarity between users on tags, and once the similarity between users on re￾sources is used for recommendation. Folkrank uses random walk techniques on the user-resource-tag (URT) graph based on the idea that popular users, resources, and tags can re￾inforce each other. These algorithms take co-occurrence of tags into account only indirectly, via the URT graph. Syme￾onidis et al. [30] employ dimensionality reduction to per￾sonalized tag recommendation. Whereas [20] operate on the URT graph directly, [30] use generalized techniques of SVD (Singular Value Decomposition) for n-dimensional tensors. The 3 dimensional tensor corresponding to the URT graph is unfolded into 3 matrices, which are reduced by means of SVD individually, and combined again to arrive at a more dense URT tensor approximating the original graph. Tag recommendation then suggests tags to users, if their weight is above some threshold. An interactive approach is pre￾sented in [14]. After the user enters a tag for a new resource, the algorithm recommends tags based on co-occurence of tags for resources which the user or others used together in the past. After each tag the user assigns or selects, the set is narrowed down to make the tags more specific. In [27], Shepitsen et al. propose a resource recommendation system based on hierarchical clustering of the tag space. The rec￾ommended resources are identified using user profiles and tag clusters to personalize the recommendation results. Us￾ing LDA topic models to recommend resources rather than tags is subject for future work. An approach to collective tag recommendation based on association rule mined from the resource tag matrix has been introduced in [18]. As discussed in Section 3.2, this ap￾proach recommends rather general tags with low TFIDF, and achieves smaller recall and precision than the approach based on LDA introduced in this paper. When content of resources is available, tag recommendation can also be ap￾proached as a classification problem, predicting tags from content. A recent approach in this direction is presented in [29]. They cluster the document-term-tag matrix after an approximate dimensionality reduction, and obtain a ranked membership of tags to clusters. Tags for new resources are recommended by classifying the resources into clusters, and ranking the cluster tags accordingly. Tags have been proven to be very useful for search: in case of image search where content based features are very difficult to extract [12], in case of enterprise search where not enough link information is available [13], or in case of web search to optimize results [3]. A large scale evaluation of Delicious regarding search is presented in [17]. They found that 50% of the pages annotated by a particular tag contain the tag within the page’s content. Bischoff et al. [9] provide an indepth analysis of a number of tagging systems with respect to to their usefulness for search. They observe that descriptive tags such as topic or type tags are much more frequent than personal tags such as ”to read”, especially in the mid and low tag frequency range, and that these tags are indeed used in search. Berendt and Hanser [6] argue that tags can be considered content and not just metadata which makes them valuable in a content based document retrieval scenario as well. Recently a number of papers deal with imroving search in tagging systems. Krestel and Chen [21] propose a method to measure the quality of tags with respect to the annotated resource to identify high quality tags that describe a re￾source better than others. Hotho et al. [19] propose exploit￾ing co-ocurrence of users, resources, and tags for searching and ranking within tagging systems. “FolkRank” is using a graph model to represent the folksonomy and can be used to rank classical keyword search results. In [5], Begelman et al. present a tag clustering algorithm to improve search. The setting is similar to ours: Related tags are identified that can be used for extending existing resource annotations, query expansion or result clustering. The clustering is based on simple co-occurence counts. Unfortunately, the paper does not contain a sound evaluation of the results. Schenkel et al. [26] propose to improve search in tagging systems by ex￾panding a user query with semantically similar tags and rank the result additionally based on a social component, which means that tagging information of friends of a user in the network is taken into account when a user submits a query