Abstract
Machine learning competitions such as those organized by Kaggle or KDDrepresent a useful benchmark for data science research. In this work, wepresent our winning solution to the Game Data Mining competition hosted at the2017 IEEE Conference on Computational Intelligence and Games (CIG 2017). Thecontest consisted of two tracks, and participants (more than 250, belonging toboth industry and academia) were to predict which players would stop playingthe game, as well as their remaining lifetime. The data were provided by amajor worldwide video game company, NCSoft, and came from their successfulmassively multiplayer online game Blade and Soul. Here, we describe the longshort-term memory approach and conditional inference survival ensemble modelthat made us win both tracks of the contest, as well as the validationprocedure that we followed in order to prevent overfitting. In particular,choosing a survival method able to deal with censored data was crucial toaccurately predict the moment in which each player would leave the game, ascensoring is inherent in churn. The selected models proved to be robust againstevolving conditions---since there was a change in the business model of thegame (from subscription-based to free-to-play) between the two sample datasetsprovided---and efficient in terms of time cost. Thanks to these features andalso to their a ability to scale to large datasets, our models could be readilyimplemented in real business settings.