where w is a weight vector with one entry for each node in V, A is the row-stochastic version of the adjacency matrix13 of the graph GF defined above, p is the random surfer vector – which we use as preference vector in our setting, and d ? [0,1] is determining the strength of the influence of p. By normalization of the vector p, we enforce the equality ||w||1 = ||p||1. This14 ensures that the weight in the system will remain constant. The rank of each node is its value in the limit w := limt?wt of the iteration process.
For a global ranking, one will choose p = 1, i. e., the vector composed by 1’s. In order to generate recommendations, however, p can be tuned by giving a higher weight to the user node and to the resource node for which one currently wants to generate a recommendation. The recommendation T? (u, r) is then the set of the top n nodes in the ranking, restricted to tags.
As the graph GF is undirected, most of the weight that went through an edge at moment t will flow back at t+1. The results are thus rather similar (but not identical, due to the random surfer) to a ranking that is simply based on edge degrees. In the experiments presented below, we will see that this version performs reasonable, but not exceptional. This is in line with our observation in [12] which showed that the topic-specific rankings are biased by the global graph structure. As a consequence, we developed in [12] the following differential approach.
FolkRank – Topic-Specific Ranking The undirectedness of graph GF makes it very difficult for other nodes than those with high edge degree to become highly ranked, no matter what the preference vector is. This problem is solved by the differential approach in FolkRank, which computes a topic-specific ranking of the elements in a folksonomy. In our case, the topic is determined by the