Finding ways of reducing undesired behavior in online interactions is at the forefront of the social computing research agenda. One promising way to reduce perceived "bad behavior" is by matching members of online social environments with corresponding behavioral preferences. We present an empirical study (N = 267) of an experimental system for matching users. As a test bed we chose the online game MechAssault, which largely supports one-off encounters within a socially homogenous population. Even within this population we found great variability in the way users selected their gaming partners. One type of player chose partners mainly on their skill, another mainly on a friendly gaming personality, and a third preferred aggressive players. Detailed analyses revealed the underlying attributes of user profiles that generated these user types. The findings suggest that a matchmaking system can better promote desired online interactions than the enforcement of uniform behavioral standards.