Dark Personality Prediction from Player in-game Behavior: Machine Learning Methods

Abstract: [Objective] By utilizing players’ behavioral data in DOTA2, this study proposed a non-intrusive method to identify the Dark Personality of game players. [Methods] After extracting behavioral features from DOTA2 replay files with the help of the parsing tool Clarity 2 package, and obtaining players’ dark personality through Dirty Dozen, we recruited machine learning methods to predict players’ sub-dimensions of Dark Personality. [Results] Results showed that best performance occurred with Gaussian Process Regression on Machiavellianism, narcissism and psychopathy. The correlations between predicting values and actual values were between 0.31 and 0.45, and the test-retest correlations were between 0.33 and 0.53. [Limitations] This study did not involve players’ verbal behavior in the process of establishing models, resulting that the features sets were not comprehensive enough. [Conclusions] It suggested that in-game behavior data was able to help predict Dark Personality of players, and the models built by Gaussian Process Regression had the best results in terms of validity and reliability.