ers being primarily located in China, playing accounts continuously until they are banned, not being skilled at non-routine in-game tasks, and avoiding the story-based elements of the game.

However, some variables which should have been obvious validity checks for whether an account was or was not involved in gold farming acted unexpectedly. Killing many NPCs actually significantly reduced the odds of being identified as a gold farmer while having a large amount of wealth in a bank did not significantly increase or decrease the odds. Farmers’ behavior evidently did not differ significantly and was likely concealed by the behavior of other players. Certainly, elite and hardcore players have killed many NPCs and accumulated great wealth. Alternatively, farming accounts that have escaped detection also exhibit this pattern. Using machine learning classifiers to predict gold farming status, the precision and recall of these models was surprisingly low. In fact, the classifier algorithms were returning many false positives—accounts fitting the profile for other gold farmers but never identified and banned by the game administrators. Clearly our reference set was not an ideal set.

SOE administrators, like many law enforcement organizations, rely on reports from players, patrols through the game world, coordinated sting operations, and database sleuthing to identify and “roll up” gold farmers. Based on the administrators’ experience, they observed farmers employing increasingly sophisticated organizations and supply chains they described like a drug trafficking operation. There are the “farmers” who actually collect the gold from the game environment and send into the distribution network, the “mules” who move this money between other agents in the network, the “dealers” who interact with customers to give them the virtual items, the short-lived “marketers” who spam chat channels, and the “wholesalers” or “bankers” who receive and distribute goods but otherwise remain inactive to avoid attention. Clearly, this was not a simple binary classification task— each of these roles had very distinct behavioral profiles.

In light of the fact there are simultaneously many gold farmer roles or classes as well as a significant number of unidentified farming accounts, we shifted the attention of our analysis [ 6]. Because trade of items and in-game currency are the fundamental operations in a gold farming network, we decided to employ a network analytic perspective. The trade network was constructed from the list of the exchanges of items and money between players by direct interaction or in-game mail. These could include legitimate bartering (trading an item for an item), market exchange (trading an item for currency), as well as unreciprocated “gifts” (sending currency or item, but receiving nothing in return).

We examined the known farmer’s
prior trade exchanges with other char-
acters that were never banned. This
naïve suspect-by-association heuristic
allowed us to define a set of farming
affiliates: these could be the paying
customers of a gold farmer, unidenti-
fied farmers, or unsuspecting players.
The majority of players who never in-
teracted with a farmer were classified
as non-affiliates. This trade network
had a number of interesting topologi-
cal properties. In keeping with the
mapping philosophy of comparing
and testing these in-game networks
to their real-world counterparts, we
examined these features both for the
game data and for a dataset of known
Canadian drug traffickers developed
by Carlo Morselli—known as the CAV-
IAR network. Given that criminal and
clandestine organizations are assem-
bled on underlying trust relationships,
net work analyses of trust proxies—like
trade and exchange relationships—
could reveal important patterns about
how individuals in these organizations
are distributed throughout the entire
network.