Database Reference
In-Depth Information
Two-mode and three-mode communities are relatively less studied mainly due to
the increased complexity and sophistication of the methods necessary for their
detection. Identifying such communities relies on bipartite and tripartite graph
clustering, which is computationally more expensive. Several simplified approaches
for deriving such communities [ 20 , 23 ] have delivered promising results.
5.2.4 Community Discovery Methods
Community identification can be thought of as a kind of clustering, i.e., a process
that splits the objects of a dataset into meaningful groups. It is for this reason that
conventional clustering techniques, such as k -means and Hierarchical Agglomer-
ative Clustering, have been used in the context of Collaborative Tagging Systems
to derive communities. For instance, in [ 9 ], a variant of the k -means algorithm is
employed to cluster a set of flickr users into groups based on their tagging
behavior. Hierarchical Agglomerative Clustering is even more popular, e.g., for
clustering tags in [ 14 , 15 ]. Recently, more sophisticated clustering schemes have
been applied on folksonomies, for example, co-clustering [ 20 ] and tensor-based
spectral clustering [ 13 ].
Despite their wide adoption, the application of conventional clustering techni-
ques is troubled by two main problems, which are especially profound in the
context of Collaborative Tagging Systems. First, clustering algorithms typically
require the number of clusters (communities) to be set a priori. Exploring the effect
of such a parameter in datasets of limited size is usually not a problem. However, in
the context of a real-world folksonomy dataset, there is no reliable method for
estimating the number of communities. Furthermore, conventional clustering algo-
rithms operate on the whole dataset in order to produce the cluster structure and
involve significant computational complexity. For this reason, the magnitude of
real-world tagging datasets renders many of these methods impractical for detecting
communities in Collaborative Tagging Systems.
The two above reasons encourage the use of community detection methods [ 27 ]
in the context of Collaborative Tagging Systems. Community detection methods are
based on a graph representation of the objects under study and exploit the graph
structure in order to group objects into communities. Community detection methods
attempt to find the “natural” organization of the objects of a system into groups.
Thus, they do not need the number of communities to be provided as parameter.
Furthermore, new community detection techniques have recently appeared that are
scalable to datasets of very large size, thus being suitable for the analysis of
community structure in real-world tagging systems.
One should note that the majority of works applying community detection on
folksonomy networks [ 5 , 12 , 28 , 29 ] has made use of modularity-optimization
schemes such as the one in [ 30 ]. Thus, such methods address the first of the above
problems: they can identify the number of communities present in the network under
study. Furthermore, there are computationally efficient schemes for optimizing
Search WWH ::




Custom Search