Bandit Algorithms for Efficient Toxicity Detection in Competitive Online Video Games

This article considers the problem of efficient sampling for toxicity detection in competitive online video games. Video game service operators take proactive measures to detect and address undesirable behavior, seeking to focus these costly efforts where such behavior is most likely. To achieve this objective, service operators need estimates of the likelihood of toxic behavior. When no pre-existing predictive model of toxic behavior is available, one must be estimated in real-time. To this end, we propose a contextual bandit algorithm that uses a small set of variables, selected based on domain expertise, to guide monitoring decisions. This algorithm balances exploration and exploitation to optimize long-term performance and is designed intentionally for easy deployment in production environments. Using data from the popular first-person action game Call of Duty (R): Modern Warfare (R) III, we show that our algorithm consistently outperforms baseline algorithms that rely solely on individual players' past behavior, achieving improvements in detection rate of up to 24.56 percentage points or 51.5%. These results have substantive implications for the nature of toxicity and illustrate how domain expertise can be harnessed to help video game service operators detect and address toxicity, ultimately fostering a safer and more enjoyable gaming experience.