Computational social science is an emerging interdisciplinary field which uses computational techniques to model, simulate and analyze human behavior and social phenomena, typically by leveraging the power of big data. Some researchers in this field use the records of actions taken by players in online games, meaning that they obtain a large amount of human behavioral data without explicitly conducting experiments with human subjects. This study investigates the mechanism of a social phenomenon called behavioral bursts which occur in a social networking game with over 7 million players. We define a behavioral burst as an increase in average points obtained by players that goes over the estimated value by a threshold. The results show that the burst patterns are classified into two classes depending on their cause. One is exogenous burst, caused by external factors (e.g. end of events), and the other is endogenous burst, caused by player interaction. We then investigate the change in the behavior of players between before and during the bursts, and find that only a small number of players are involved in them in case of endogenous bursts. Furthermore, by focusing on the action sequences of players obtaining many points, we find that they communicate with other players in a modest number of times, and in the case of endogenous bursts they tend to exchange information with a fixed player.