C. Wartena and m. wibbels of different algorithms for top-n recommendation topic aware by clustering the tags or items in the profile and generating separate recommendation lists for each topic cluster. The final list of recommendations is obtained by merging the topic specific lists. We do not consider rating prediction. Lists of recommended items are more interesting to a user if they are more di verse. Thus diversity is regarded as a desirable property of recommendations. In most studies on topic diversification, the diversity is increased by reordering the elements in an initial list of recommended items. If a list is constructed aiming at optimal results for precision and recall, the reordering usually causes a decrease of performance on these evaluation measures. Thus a trade-off between accuracy and diversity emerges In our approach, however, adding diversity improves pre cision and recall. We do not re-rank results that are already optimal for precision and recall, but the final diversity of the recommendation is a core property of the recommendation strategy. Thus our method is fundamentally different from the re-ranking approaches: we first distinguish different topics and then generate a list of relevant items The remainder of this paper is organized as follows. In the next section we discuss related work. In section 3 we introduce three different tag based recom- mendation algorithms In section 4 we discuss topic detection for tagging system and define topic aware versions of the tag based algorithms In section 5 we port on an evaluation of the proposed algorithms with data from Library Thing, a bookmarking service for book 2 Related work Most work on recommendation and tagging is about recommend ing tags for item prediction has received less attention. Basically, we find two approaches for tag-based item recommendation. The first approach uses tags t compute item-item or user-user similarities that then are used in classical user or item based nearest neighbor recommendation algorithms. The second approach characterizes both users and items by tag vectors, making it possible to compute the similarity between users and items. The items that are most similar to a user now are recommended One of the first papers that integrates tag-based similarities in a nearest neigh bors recommender is [1, who extend the user-item matrix with user-tag simi- larities in order to compute user-user similarities, and extend it with tag-item relations in order to compute item-item similarities. Both similarities are used to compute recommendations. This approach was refined by [2 taking also into account whether users used the tags for the same or for different items. Said et al. 3 use probabilistic latent semantic analysis to model both the user-item matrix and the item-tag matrix. By sharing the latent variables between the modelsthey are able to use the tagging information in the user-item model. Bogers and Van den Bosch [4 use the tag based similarities instead of the clas- sical similarities based on the user-item matrix. They show improvements for item-based nearest neighbor recommendation on various datasets, but did not compare their method to approaches combining both types of similarity44 C. Wartena and M. Wibbels of different algorithms for top-n recommendation topic aware by clustering the tags or items in the profile and generating separate recommendation lists for each topic cluster. The final list of recommendations is obtained by merging the topic specific lists. We do not consider rating prediction. Lists of recommended items are more interesting to a user if they are more di￾verse. Thus diversity is regarded as a desirable property of recommendations. In most studies on topic diversification, the diversity is increased by reordering the elements in an initial list of recommended items. If a list is constructed aiming at optimal results for precision and recall, the reordering usually causes a decrease of performance on these evaluation measures. Thus a trade-off between accuracy and diversity emerges. In our approach, however, adding diversity improves pre￾cision and recall. We do not re-rank results that are already optimal for precision and recall, but the final diversity of the recommendation is a core property of the recommendation strategy. Thus our method is fundamentally different from the re-ranking approaches: we first distinguish different topics and then generate a list of relevant items. The remainder of this paper is organized as follows. In the next section we discuss related work. In section 3 we introduce three different tag based recom￾mendation algorithms. In section 4 we discuss topic detection for tagging systems and define topic aware versions of the tag based algorithms. In section 5 we re￾port on an evaluation of the proposed algorithms with data from LibraryThing, a bookmarking service for books. 2 Related Work Most work on recommendation and tagging is about recommending tags. Us￾ing tags for item prediction has received less attention. Basically, we find two approaches for tag-based item recommendation. The first approach uses tags to compute item-item or user-user similarities that then are used in classical user or item based nearest neighbor recommendation algorithms. The second approach characterizes both users and items by tag vectors, making it possible to compute the similarity between users and items. The items that are most similar to a user now are recommended. One of the first papers that integrates tag-based similarities in a nearest neigh￾bors recommender is [1], who extend the user-item matrix with user-tag simi￾larities in order to compute user-user similarities, and extend it with tag-item relations in order to compute item-item similarities. Both similarities are used to compute recommendations. This approach was refined by [2] taking also into account whether users used the tags for the same or for different items. Said et al. [3] use probabilistic latent semantic analysis to model both the user-item matrix and the item-tag matrix. By sharing the latent variables between the models they are able to use the tagging information in the user-item model. Bogers and Van den Bosch [4] use the tag based similarities instead of the clas￾sical similarities based on the user-item matrix. They show improvements for item-based nearest neighbor recommendation on various datasets, but did not compare their method to approaches combining both types of similarity