A ranked list of entities is assumed to
positively improve the selection of the most
relevant candidate cluster (the ranking algorithm is
based on AktiveRank by Alani et al. [2006] and is
thoroughly described in [Tomassen & Strasunskas,
2009]). An underlying assumption is that more
information is available for the most central
entities (they are the most semantically rich by
having the largest number of relations and by
being central to other entities) and consequently
they are better candidates to distinguish relevant candidate clusters. The most prominent cluster
candidates are later used to identify new candidate
clusters, and so on. However, a potential problem
with this approach is the drifting of focus (i.e., an
erroneous candidate cluster is selected which next
is used to find the most prominent candidate
cluster of a neighbouring entity, and so forth).