Improving Tag-Based Recommendation by Topic Diversification The second approach, recommendation based on the distance between an item and the tag based user profile, is e. g. followed by Firan et al. 5. The focus of their work is to determine the optimal set of tags to represent a user. The obvious idea is to represent a user by the tags that he has assigned. However, the resulting tag vector is usually too sparse to compute useful user-item similarities. Some users employ only a very small set of tags, and even for more actively tagging users it might be well the case that a relevant item is tagged only with synonyms of a tag employed by the user. Thus 5 investigate various methods to condense the user profile. The most effective method is to use the tags of all items the user has bookmarked. The same observation was also made by [6. The problem of the sparse user profiles was also identified by [7. They solve this problem not by condensing the user profile, but by taking co-occurring tags into account in the computation of similarities. For each user and each tag t a user specific tag weight is computed. The weight for t is determined by the weight of the most similar tag t' in the user profile and the similarity between t and t', where the inter tag similarity is determined by a variant of the Jaccard-coefficient. The relevance of an item finally is the sum of weights of its tags In 8 the two approaches sketched above are combined: a nearest neighbor gorithm is used to find an initial set of items and subsequently user-item simi- larities are computed for the preselected items to obtain a final recommendation The more interesting aspect of their approach is, that they replace each tag t in the user profile with the tags co-occurring with t on items tagged by the user and restrict the set of tags to the (globally) most popular tags. This results in roughly the same profiles as would be obtained by using the(most popular) tags of all items bookmarked (or tagged) by the user, as we have seen in other approaches. The problem that many recommender systems tend to recommend a set of rery similar items in order to optimize accuracy was noted by 9 and coined the portfolio effect. Reordering of recommended elements to alleviate this problem was proposed by [10. As discussed above, the reordering increases diversity, but at the cost of accuracy. An approach that is very similar to ours is proposed by Zhang and Hurley [11]. They cluster the set of items of a user and apply a item based nearest neighbor recommender to each of the clusters. Finally they merge the results of the sub-recommendation to obtain a list of recommended popular items is recommended very often at the cost of novel or less common items. The main difference with our recommendation strategy is that all item similarities are based on the user-item matrix, whereas we base similarities on descriptive meta data, especially tags. Zhang and Hurley can improve diversity of the recommendation lists in some cases while the influence of the partitioning on the precision is not very large Gemmell et al. [12 propose to cluster tags in a user profile to improve person lized search. They show that clustering improves the search results. Gemmell et al did not consider recommendation, but our approach for recommendation is both in spirit and results similar to their work. An interesting solution targetingImproving Tag-Based Recommendation by Topic Diversification 45 The second approach, recommendation based on the distance between an item and the tag based user profile, is e.g. followed by Firan et al. [5]. The focus of their work is to determine the optimal set of tags to represent a user. The obvious idea is to represent a user by the tags that he has assigned. However, the resulting tag vector is usually too sparse to compute useful user-item similarities. Some users employ only a very small set of tags, and even for more actively tagging users it might be well the case that a relevant item is tagged only with synonyms of a tag employed by the user. Thus [5] investigate various methods to condense the user profile. The most effective method is to use the tags of all items the user has bookmarked. The same observation was also made by [6]. The problem of the sparse user profiles was also identified by [7]. They solve this problem not by condensing the user profile, but by taking co-occurring tags into account in the computation of similarities. For each user and each tag t a user specific tag weight is computed. The weight for t is determined by the weight of the most similar tag t in the user profile and the similarity between t and t , where the inter tag similarity is determined by a variant of the Jaccard-coefficient. The relevance of an item finally is the sum of weights of its tags. In [8] the two approaches sketched above are combined: a nearest neighbor algorithm is used to find an initial set of items and subsequently user-item simi￾larities are computed for the preselected items to obtain a final recommendation. The more interesting aspect of their approach is, that they replace each tag t in the user profile with the tags co-occurring with t on items tagged by the user, and restrict the set of tags to the (globally) most popular tags. This results in roughly the same profiles as would be obtained by using the (most popular) tags of all items bookmarked (or tagged) by the user, as we have seen in other approaches. The problem that many recommender systems tend to recommend a set of very similar items in order to optimize accuracy was noted by [9] and coined the portfolio effect. Reordering of recommended elements to alleviate this problem was proposed by [10]. As discussed above, the reordering increases diversity, but at the cost of accuracy. An approach that is very similar to ours is proposed by Zhang and Hurley [11]. They cluster the set of items of a user and apply a item based nearest neighbor recommender to each of the clusters. Finally they merge the results of the sub-recommendation to obtain a list of recommended items for the user. The focus of their work is to avoid that a small set of very popular items is recommended very often at the cost of novel or less common items. The main difference with our recommendation strategy is that all item similarities are based on the user-item matrix, whereas we base similarities on descriptive meta data, especially tags. Zhang and Hurley can improve diversity of the recommendation lists in some cases while the influence of the partitioning on the precision is not very large. Gemmell et al. [12] propose to cluster tags in a user profile to improve person￾alized search. They show that clustering improves the search results. Gemmell et al. did not consider recommendation, but our approach for recommendation is both in spirit and results similar to their work. An interesting solution targeting