Open world video games are designed to offer free-roaming virtual environments and agency to the players, providing a substantial degree of freedom to play the games in the way the individual player prefers. Open world games are typically either persistent, or for single-player versions semi-persistent, meaning that they can be played for long periods of time and generate substantial volumes and variety of user telemetry. Combined, these factors can make it challenging to develop insights about player behavior to inform design and live operations in open world games. Predicting the behavior of players is an important analytical tool for understanding how a game is being played and understand why players depart (churn). In this paper, we discuss a novel method of learning compressed temporal and behavioral features to predict players that are likely to churn or to continue engaging with the game. We have adopted the Relaxed Tensor Dual DEDICOM (RTDD) algorithm for bipartite tensor factorization of temporal and behavioral data, allowing for automatic representation learning and dimensionality reduction.