Social bookmarking systems have established themselves as an important part in today's Web. In such systems, tag recommender systems support users during the posting of a resource by suggesting suitable tags. Tag recommender algorithms have often been evaluated in offline benchmarking experiments. Yet, the particular setup of such experiments has rarely been analyzed. In particular, since the recommendation quality usually suffers from difficulties such as the sparsity of the data or the cold-start problem for new resources or users, datasets have often been pruned to so-called cores (specific subsets of the original datasets), without much consideration of the implications on the benchmarking results. In this article, we generalize the notion of a core by introducing the new notion of a set-core, which is independent of any graph structure, to overcome a structural drawback in the previous constructions of cores on tagging data. We show that problems caused by some types of cores can be eliminated using set-cores. Further, we present a thorough analysis of tag recommender benchmarking setups using cores. To that end, we conduct a large-scale experiment on four real-world datasets, in which we analyze the influence of different cores on the evaluation of recommendation algorithms. We can show that the results of the comparison of different recommendation approaches depends on the selection of core type and level. For the benchmarking of tag recommender algorithms, our results suggest that the evaluation must be set up more carefully and should not be based on one arbitrarily chosen core type and level.
