Identification and Classification of Player Types in Massive Multiplayer Online Games Using Avatar Behavior

TLDR: An improved methodology for classifying players (identifying deviant players such as terrorists) through multivariate analysis of data from avatar characteristics and behaviors in massive multiplayer online games (MMOGs) is developed.

Abstract: : The purpose of our research is to develop an improved methodology for classifying players (identifying deviant players such as terrorists) through multivariate analysis of data from avatar characteristics and behaviors in massive multiplayer online games (MMOGs). To build our classification models, we developed three significant enhancements to the standard Generalized Regression Neural Networks (GRNN) modeling method. The first enhancement is a feature selection technique based on GRNNs, allowing us to tailor our feature set to be best modeled by GRNNs. The second enhancement is a hybrid GRNN which allows each feature to be modeled by a GRNN tailored to its data type. The third enhancement is a spread estimation technique for large data sets that is faster than exhaustive searches, yet more accurate than a standard heuristic. We applied our new techniques to a set of data from the MMOG, Everquest II, to identify deviant players ('gold farmers'). The identification of gold farmers is similar to labeling terrorists in that the ratio of gold farmer to standard player is extremely small, and the in-game behaviors for a gold farmer have detectable differences from a standard player. Our results were promising given the difficulty of the classification process, primarily the extremely unbalanced data set with a small number of observations from the class of interest. As a screening tool our method identifies a significantly reduced set of avatars and associated players with a much improved probability of containing a number of players displaying deviant behaviors. With further efforts at improving computing efficiencies to allow inclusion of additional features and observations with our framework, we expect even better results.