Improving Tag-Based Recommendation by Topic Diversification 3.1 Item Based Collaborative Filtering The first strategy for recommendation we use is a nearest neighbor approach (13 ). Given a distance measure between items, we express the relevance of an item for a given user as the average distance between the item and all items bookmarked (or seen) by the user. Formally, we define the distance of an item i∈ C to a user u∈Uas where d(i, j)is the distance between items i and j. The items with the smallest distance are recommended to the user Given our perspective of tag distributions it is natural to use divergences for e distance between items. We will base our distance on the ensen Shannon divergence, which can be considered as a symmetrical variant of the Kullback Leibler divergence or relative entropy. Since the square root of the Jensen Shan- non divergence is a proper metric satisfying the usual axioms of non-negativity, identity of indiscernibles and triangle inequality(14), we will use d(i,j)=VJSD (PT(ti), Pr(+)) where JSD(p, q) is the Jensen Shannon divergence or information radius between probability distributions p and g. 3.2 Profile based recommendation In the nearest neighbor approach we compute the average distance of an item to the items seen by the user. Alternatively, we can compute the distance of an item to the user that also can be represented by a distribution over tags. For each user we compute the characteristic tag distribution p(tu). The obvious way to define this distribution is: Pr(tu)=n(u, t)/nu(u) (10) This allows us to define a distance between an item i and a user u by setting d(u,) D(pT(tu),pr(t)) and to recommend items with a small distance to the user. However, it was already shown by 5 that this strategy does not perform very well In our experiments the results were even worse than those btained by the simple non-personalized baseline of recommending the most popular items. The main reason for this bad performance is that the distribution is too sparse and thus many relevant items will be tagged with synonyms of the tags in the user profile, but not with exactly those tags In the following we will present two possibilities to alleviate this problem Profiles based on item tags. Firan et al.(5) propose several methods to construct better user profiles. One of the most successful ones is to use the tags of the items considered by the user. We define the item based user profile as pr(tu) PTImproving Tag-Based Recommendation by Topic Diversification 47 3.1 Item Based Collaborative Filtering The first strategy for recommendation we use is a nearest neighbor approach ([13]). Given a distance measure between items, we express the relevance of an item for a given user as the average distance between the item and all items bookmarked (or seen) by the user. Formally, we define the distance of an item i ∈ C to a user u ∈ U as d(u, i) = j∈Cu d(j, i) |Cu| (8) where d(i, j) is the distance between items i and j. The items with the smallest distance are recommended to the user. Given our perspective of tag distributions it is natural to use divergences for the distance between items. We will base our distance on the Jensen Shannon divergence, which can be considered as a symmetrical variant of the Kullback Leibler divergence or relative entropy. Since the square root of the Jensen Shan￾non divergence is a proper metric satisfying the usual axioms of non-negativity, identity of indiscernibles and triangle inequality ([14]), we will use d(i, j) = JSD (pT (t|i), pT (t|j)) (9) where JSD(p, q) is the Jensen Shannon divergence or information radius between probability distributions p and q. 3.2 Profile Based Recommendation In the nearest neighbor approach we compute the average distance of an item to the items seen by the user. Alternatively, we can compute the distance of an item to the user that also can be represented by a distribution over tags. For each user we compute the characteristic tag distribution p(t|u). The obvious way to define this distribution is: pT (t|u) = n(u, t)/nU (u). (10) This allows us to define a distance between an item i and a user u by setting d(u, i) = JSD (pT (t|u), pT (t|i)) and to recommend items with a small distance to the user. However, it was already shown by [5] that this strategy does not perform very well. In our experiments the results were even worse than those obtained by the simple non-personalized baseline of recommending the most popular items. The main reason for this bad performance is that the distribution is too sparse and thus many relevant items will be tagged with synonyms of the tags in the user profile, but not with exactly those tags. In the following we will present two possibilities to alleviate this problem. Profiles based on item tags. Firan et al. ([5]) propose several methods to construct better user profiles. One of the most successful ones is to use the tags of the items considered by the user. We define the item based user profile as p T (t|u) = 1 |Cu| i∈Cu pT (t|i). (11)