19.5 Comparison of Algorithms
In this section, we briefly discuss the main advantages and disadvantages of the algorithms presented in Section 19.4. Note that we just consider the non contentbased algorithms since they can be compared under a common basis.
We saw in Section 19.4.1 that in order to apply standard CF-based algorithms to folksonomies, some data transformation must be performed. Such transformations lead to information loss, which can lower the recommendation quality. Another well known problem with CF-based methods is that large projection matrices must be kept in memory, which can be time/space consuming thus compromising the ability to perform real-time recommendations. Another problem is that for each different mode to be recommended, the algorithm must be eventually changed, demanding an additional effort for offering multi-mode recommendations.
FolkRank builds on PageRank and proved to give significantly better tag recommendations than CF [15]. This method also allows for mode switching with no change in the algorithm. Moreover, as well as CF-based algorithms, FolkRank is robust against online updates since it does not need to be trained every time a new user, resource or tag enters the system. However, FolkRank is computationally expensive and not trivially scalable, making it more suitable for systems where real-time recommendations is not a requirement.
Similarly to FolkRank, tensor factorization methods work directly on the ternary relations of folksonomies. Although the learning phase can be costly, it can be performed offline. After the model is learned, the recommendations can be done fast, making these algorithms suitable for real-time recommendations. A potential disadvantage of tensor factorization methods is that easy mode switching can only be achieved if one consider that the different recommendation problems, i.e., user/resource/tag, can be addressed by minimizing the same error function. If one chooses HOSVD, for example, the model can be used for multi-mode recommendations with trivial mode switching, but at the cost of evtl. solving the wrong problem: HOSVD minimizes a least-square error function while social tagging RS are more related to ranking. If one tries to optimally reconstruct the tensor w.r.t. an error function targeted to a specific recommendation mode, on the other hand, accuracy is eventually improved for the targetted mode, but at the cost of making mode switching more involved. Figure 19.9 shows a comparison between some of the aforementioned algorithms in a snapshot of the BibSonomy dataset for the tag recommendation problem [26]. Note that the best method is RTF followed by FolkRank and HOSVD.
Table 19.2 summarizes this discussion. Note that the absence of a “X” in Table 19.2 indicates that the corresponding property is not trivially achieved by the algorithm being considered.