Information Technology Reference
In-Depth Information
run. This gave us a good insight about how each interaction strategy e.g., 'random'
meeting, 'friend-of-a-friend', 'party' affect the above global measures of the simu-
lated social network. Based on these measures and exploration techniques for parame-
ter space, we identified one set of the interaction model as the best candidate to
explain the underlying reference dataset.
To initiate the debate on the need for standards to validate simulated networks, we
suggest to begin with calculating topological measures such as degree distribution,
community structure, clustering coefficient, geodesic measures such as diameter,
radius and average path lengths and then compare them with reference dataset using
the techniques we described earlier. For social networks with nodal attributes, either
categorical or not, we propose to first calculate affinity measures for each attribute,
and then calculate, for each attribute, Silo indices for the whole attribute space. Final-
ly, perform statistical tests, if possible nonparametric; to test how best they fit the
data. These set of measures develop confidence in validating our models and could be
deemed as a standard set of measures for the social network comparison.
8
Concluding Discussion
There does not seem to be a “golden bullet” for comparing simulation-generated net-
works and empirically derived networks. This paper provides a brief review of some
of the approaches that are in contemporary use. These approaches, which are by no
means exhaustive, could play some part in determining the extent to which a generat-
ed graph matches a target graph. Clearly, the more complex a model is (and agent-
based simulations tend to be at the more complex end of the modeling spectrum), the
more independent validation it needs, suggesting that a multiple approach is desirable.
The kind of network validation attempted should depend upon the model's underlying
assumptions and goals behind the simulation - namely, what is and is not, deemed
significant about the reference and simulated networks. Ideally, these should be do-
cumented in any description of the model validation so a reader is in a position to
judge the claimed goodness of the simulation network “fit”. There does not seem
to be an established norm for either describing such underlying assumptions nor
for measuring the extent to which a particular set of generated networks match their
targets.
A majority of agent-based social simulation models uses stereotypic networks
(e.g., the Watts-Strogatz network), over which agents interact. A smaller set of mod-
els aim at simulating social networks through more descriptive and contextualized
rules (e.g., [4]) but these often lack the data to sufficiently validate the resulting net-
works. The recent rise of online social networks offers the possibility to acquire the
necessary data for the network validation. For instance, Abbas [1, 2, 3] used the
friendship networks for a college campus in the United States to help validate an
agent-based model of friendship choice. Here networks generated from different set
of rules from the agent-based model were compared using several of the network
measures above.
