Graphics Reference
In-Depth Information
the loosest of definitions, it may be better in some cases to identify only
three.
•
Factoring in link weight
—In most cases stronger links between
nodes should be interpreted as tighter community bonds. This can be
factored into computation. In technical terms, link strength in graphs is
often referred to as
edge weight
. Links may, for example, be weighted
by the number of transactions between two nodes or by the total value
of all transactions if they vary in size.
•
Consideration of node similarity
—In some cases, it is useful to
identify communities not only by how tightly they are linked but also by
factoring in similarity of node attributes (such as demographic profile).
For example, a cluster of closely associated, mostly middle-income Baby
Boomers may be more actionable from a marketing perspective than
simply a cluster of tightly linked customer accounts of all varieties.
•
Overlapping or discrete clusters
—In reality, communities often
overlap. However, for the purposes of analysis and decision making, it
can sometimes be more useful if each node is assigned to only one
cluster. In such cases, it is often common practice to assign all
remaining nodes to an additional cluster representing “other.”
•
Compute time
—Some algorithms are inherently more time
consuming to calculate than others. Depending on the size of the data
set, a trade-off between time and quality may be required. Some
algorithms are designed for this and work on a progressive refinement
basis, which the user can monitor and stop when he or she achieves a
satisfactory quality of result.
You can download many experimental community detection algorithms as
plug-ins in graph tools like Gephi or Cytoscape. The Louvain modularity
algorithm, however, provides one of the best balances of general utility
and performance, making it a good choice as the core implementation of
community detection in Gephi. Louvain strengths include relatively simple
control over the number of communities, consideration of link weight,
discrete classification of all nodes, and relatively fast computation.
In the Overview task mode, select the Statistics tab and click the Run button
beside the Modularity statistic to execute the Louvain algorithm. Leave the
option to use weights checked to consider strength of connection between
members. You can use the resolution factor to control the number of
