Improving Tag-Based Recommendation by Topic Diversification algorithms we also compare precision and diversity of a top 10 recommendation As a measure of diversity we take the average squared distance between the recommended items. The diversity for a set items I is thus defined as diu(n)>iie JSD(a(), q()) 2 where JSD is the Jensen-Shannon divergence, which is the square of a proper distance measure. The results for MAP, precision at 10 and diversity at 10 are given in Table 1 +Co-occur. tag 一 Co-occur.tags(TA) tr Nearest neighbors cHIte tags Fig 1. Precision and recall for 3 proposed topic aware algorithms compared to their basic (non-topic aware)variants, using a sample of Library Thing data. Data points are plotted for top-n recommendation with n= 1, 5, 10, 20... 100. In Figure 1 precision and recall for the 6 tag based algorithms are show Comparison of the algorithms clearly shows that each of the 3 discussed basic algorithms benefits from making them topic aware in the proposed way. only for a top l and a top 2 the topic aware variants of the tag based collaborative filtering and the algorithm using a profile based on co-occurring tags have a lower precision than their non-topic aware variants. In these cases the number of clusters is larger than the number of items that we have to predict. Thus, most of the available information is not used. Remarkably, the third topic aware dgorithm does not show a similar behavior for small recommendation lists If we compare the algorithms to the baselines( Figure 2)we see that all al gorithms are clearly better than the non-personalized most viewed recommen- dation. T-tests show that differences between each tag aware algorithm and its base variant for the MAP and for prec@10 are significant at the p<0, 001 level Also all differences to the base-line algorithms are significant at the same levelImproving Tag-Based Recommendation by Topic Diversification 51 algorithms we also compare precision and diversity of a top 10 recommendation. As a measure of diversity we take the average squared distance between the recommended items. The diversity for a set items I is thus defined as div(I) = i,j∈I JSD(q(i), q(j)) |I| 2 (17) where JSD is the Jensen-Shannon divergence, which is the square of a proper distance measure. The results for MAP, precision at 10 and diversity at 10 are given in Table 1. Fig. 1. Precision and recall for 3 proposed topic aware algorithms compared to their basic (non-topic aware) variants, using a sample of LibraryThing data. Data points are plotted for top-n recommendation with n = 1, 5, 10, 20 ... 100. In Figure 1 precision and recall for the 6 tag based algorithms are shown. Comparison of the algorithms clearly shows that each of the 3 discussed basic algorithms benefits from making them topic aware in the proposed way. Only for a top 1 and a top 2 the topic aware variants of the tag based collaborative filtering and the algorithm using a profile based on co-occurring tags have a lower precision than their non-topic aware variants. In these cases the number of clusters is larger than the number of items that we have to predict. Thus, most of the available information is not used. Remarkably, the third topic aware algorithm does not show a similar behavior for small recommendation lists. If we compare the algorithms to the baselines (Figure 2) we see that all al￾gorithms are clearly better than the non-personalized most viewed recommen￾dation. T-tests show that differences between each tag aware algorithm and its base variant for the MAP and for prec@10 are significant at the p < 0, 001 level. Also all differences to the base-line algorithms are significant at the same level.