Community Detection in Collaborative Tagging Systems - Community-Built Databases: Research and Development - page 115

interested at all in resources tagged with the latter. Employing tag communities as a
vector space for the user profile alleviates the above problem and thus leads to
increased recall performance for personalized retrieval systems.
Moreover, the discovery of tag communities can serve in two additional tasks:
(a) sense disambiguation [ 5 ] and (b) ontology evolution/population [ 18 ]. The work in
[ 5 ] resulted in the conclusion that clustering tag networks can lead to the identification
of the different contexts/senses that a tag is used in a potentially more effective way
than by resorting to some external source of knowledge (e.g., WordNet). In addition,
mapping tags to formal ontology concepts and tag co-occurrences to ontology rela-
tions has also been shown to be a task benefiting from the use of tag clustering
[ 18 ]. Such results demonstrate that tag community structure can be exploited as a
step toward semantifying the content in large-scale repositories containing tagged
resources.
5.3 Detection and Evaluation of Communities
in Tag Networks
From the above, it has become evident that identifying tag communities in Collab-
orative Tagging Systems is an important research problem with potential benefits
for several applications. The goal of tag community detection is to derive groups of
tags that either are semantically close to each other or share some usage context. In
practice, one may expect that the tag communities of a Collaborative Tagging
System correspond to the topics that are of interest to its users. Thus, tag commu-
nities act as a proxy of user interests at a higher level of abstraction than individual
tags and for that reason they present a more informative and context-rich means of
describing both the resources and the users of a Collaborative Tagging System.
Most existing efforts on the analysis of tag communities have employed either
conventional clustering schemes (e.g., hierarchical agglomerative clustering was
used in [ 14 , 15 ]) or the recently popularized modularity optimization schemes (used
in [ 5 , 12 , 28 , 29 ]). However, there are a number of issues specific to tag networks
that these methods do not adequately address.
Overlapping community structure . Tag communities are expected to overlap
with each other since there are numerous polysemous tags. For instance, the tag
“opera” is expected to belong to at least two communities: one related to music and
one related to browsers. Moreover, tag community overlap is expected to be more
pronounced by tags that are used in a different context by different groups of users.
For example, the tag “Barcelona” may be used by a group of people to refer to
architecture-related resources and by another group of people to refer to the city as a
travel destination. With this consideration in mind, one may expect that clustering
or community detection methods that are inherently partitional, i.e., they do not
allow for overlaps among communities, are bound to miss important information
stemming from the different contexts of tag usage.