Keep You from Leaving: Churn Prediction in Online Games

Customer retention is a crucial problem for game companies since the revenue is heavily influenced by the size of their user bases. Previous studies have reached a consensus that the cost of attracting a new player can be six times than retaining the players, which indicates an accurate churn prediction model is essential and critical for the strategy making of customer retention. Existing works more focus on studying login information (e.g. login activity traits of users) ignoring the rich in-game behaviors (e.g. upgrading, trading supplies) which could implicitly reflect user's preference from their inter-dependencies. In this paper, we propose a novel end-to-end neural network, named ChumPred, for churn prediction problem. In particular, we not only consider the login behaviors but also in-game behaviors to model user behavior patterns more comprehensively. For time series of login activities, we leverage a LSTM-based structure to learn intrinsic temporal dependencies so as to capture the evolution of activity sequences. For in-game behaviors, we develop a time-aware filtering component to better distinguish the behavior patterns occurring in a specific period and a multi-view mechanism to automatically extract the multiple combinations of these behaviors from various perspectives. Comprehensive experiments conducted on real-world data demonstrate the effectiveness of the proposed model compared with state-of-the-art methods.