C. Wartena and m. wibbels Most viewed NEarest neighbors D,1 Recall 0,15 Fig. 2. Precision and recall for 3 proposed topic aware algorithms compared to 2 base lines, using a sample of Library Thing data. Data points are plotted for top-n recom- undation with n=1. 5. 10.20....10 The results clearly show that tags can be very useful for recommendation nally, we see that the diversity for the topic aware algorithms is clearly higher than for the non-topic aware content based algorithms. Interestingly, also the diversity of BPR-MF is relatively high 6 Conclusion and Future work We have shown that clustering user tags can significantly improve tag based item recommendation. This corresponds to the intuition that people have dif- ferent interests and that it is better to recommend items on each separate topic than trying to find items that match more or less all interests. Though this is very intuitive, it is nevertheless a surprising result. Given results from previ- ous research, we expect that improvement of diversity has to be paid for by loss of precision and recall Our clustering approach in contrary improves both diversity and precision and recall Another nice aspect of the proposed algorithms is that it is easy to explain to users why items are recommended: The topics detected can be displayed, e. g by a cloud of most frequent or most central tags. The recommendation then can e motivated by the relevance of the items for one of the detected topics The two algorithms that turned out to be the best ones, do not or almost not rely on the fact that the user is an active tagger. Thus these methods can be applied for content based recommendation in general. An interesting question for future research is, to what type of item sets and meta-data the results carry52 C. Wartena and M. Wibbels Fig. 2. Precision and recall for 3 proposed topic aware algorithms compared to 2 base lines, using a sample of LibraryThing data. Data points are plotted for top-n recom￾mendation with n = 1, 5, 10, 20,... 100. The results clearly show that tags can be very useful for recommendation. Finally, we see that the diversity for the topic aware algorithms is clearly higher than for the non-topic aware content based algorithms. Interestingly, also the diversity of BPR-MF is relatively high. 6 Conclusion and Future Work We have shown that clustering user tags can significantly improve tag based item recommendation. This corresponds to the intuition that people have dif￾ferent interests and that it is better to recommend items on each separate topic than trying to find items that match more or less all interests. Though this is very intuitive, it is nevertheless a surprising result. Given results from previ￾ous research, we expect that improvement of diversity has to be paid for by loss of precision and recall. Our clustering approach in contrary improves both, diversity and precision and recall. Another nice aspect of the proposed algorithms is that it is easy to explain to users why items are recommended: The topics detected can be displayed, e.g. by a cloud of most frequent or most central tags. The recommendation then can be motivated by the relevance of the items for one of the detected topics. The two algorithms that turned out to be the best ones, do not or almost not rely on the fact that the user is an active tagger. Thus these methods can be applied for content based recommendation in general. An interesting question for future research is, to what type of item sets and meta-data the results carry over