“Bad Vibrations”: Sensing Toxicity From In-Game Audio Features

Abstract

Toxicity in online gaming is a problem that causes harm to players, developers, and gaming communities. Toxic behaviours persist in online multiplayer games for a number of reasons, and continue to go unchecked due in large part to a lack of reliable methods to accurately detect toxicity online, in real-time, and at scale. In this paper, we present a modeling approach that uses features derived from in-game verbal communication and game metadata to predict if Overwatch games are toxic. With logistic regression models, we achieve accuracy scores of 86.3% for binary (high vs low toxicity) predictions. We discuss which features were most salient, potential application of our predictive model, and implications for toxicity detection in games. Our approach is a low-cost, low-effort, and non-invasive detection approach that contributes to holistic efforts in combating toxicity in games.