rewired—a strong possibility in GSJ and the Dark Web. However, such rewiring would not affect Meth World or the gang network because a co-occurrence link could not be rewired once it was created.

An interesting topology-related question is what mechanisms play a role in producing the properties we observed in dark networks? Short average path length, high clustering coefficient, power-law degree distributions with two-regime scaling behavior in the human networks? That is, can we regenerate the four dark networks based on known mechanisms (such as growth and preferential attachment)? To answer, we conducted a series of simulations in which we generated 30 networks for each elicited human network based on three evolutionary mechanisms:

Growth. Starting with a small number of nodes, at each time step we add a new node to connect with existing nodes in the network;

Preferential attachment. The probability that an existing node will receive a link from the new node depends on the number of links the node already maintains. The more links it has the more likely it will receive a new link; and

New links among existing nodes. At each time step, a random pair of existing nodes may connect, depending on the number of common neighbors they have. The more common neighbors they share the more likely they will also be connected with each other.

We expected that the first two mechanisms would generate a power-law degree distribution¹ and that the third would generate a high clustering coefficient and two-regime scaling behavior. ² Our simulations showed that the power -law degree distributions are easily regenerated, with R² ranging from 0.83 (the gang network) to 0.88 (GSJ). The two-regime scaling behavior was also present in the simulated networks for the human networks. However, the highest clustering coefficient in a simulation was only 0.24 (GSJ), far less than what we obtained from the elicited networks (0.55–0.68). This finding implies that some other mechanisms must have contributed to the substantially high clustering coefficients we observed in the dark

the terrorists in
the GsJ network
are on average
only 2. 5 links
away from bin
Laden himself,
meaning his
command
is able to reach
an arbitrary
member through
only two mediators.

networks. We suspect that member recruitment is one such mechanism. Employing active recruitment methods, subgroups of terrorists or criminals are able to attract new members into their groups. The new members quickly become acquainted with many existing members, substantially increasing the clustering coefficients.

caveats

A notable point is that two problems may have affected the structures of the three elicited human networks—GSJ, Meth World, and the gang network. First, they may have missing links that can cause the networks to appear to be less efficient; there may actually be hidden “shortcuts” connecting distant parts of the networks. Second, the presence of coincidental “fake” links might cause the elicited networks to be more efficient than they would otherwise be since these links are not communication channels.

To test how the results would be affected by missing links, we added various percentages of the existing links to the elicited networks based on three effects used in missing-link-prediction research: ⁷

Random effect. A link is added between a randomly selected pair of nodes not originally connected;

Common neighbor effect. A link is added between a pair of unconnected nodes if they share common neighbors; the more common neighbors they share the more likely they will be connected; and

Preferential attachment effect. The probability that a pair of unconnected nodes will be linked together depends on the product of their link degrees.

We found that the small-world and scale-free properties of the four networks do not change when missing links are added. For example, when we added up to 10% of the links, the average path lengths ranged from 3. 55 links (GSJ, preferential-attachment links added) to 9. 45 links (the gang network, common-neighbor links added); the clustering coefficients ranged from 0.45 (GSJ, random links added) to 0.67 (the gang network, common-neighbor links added); and the R2 of power-law degree distributions ranged from 0.61 (GSJ, random links added) to 0.93 (the gang network,