Predictive Analytics of In-Game Transactions: Tokenized Player History and Self-Attention Techniques

Players’ purchases in free-to-play online games often serve as crucial indicators of user engagement and behavior. Understanding these purchases not only enhances the personalization of the gaming experience but also enables the optimization of game monetization strategies. This paper introduces a methodology for predicting players’ purchases using Transformers neural networks based on the Self-Attention technique, customized for processing sequential data. By discretizing the values of features representing a player’s history and leveraging tokenized inputs related to the discretized history, the methodology aims to forecast whether a player will make a purchase within the next 3, 5, or 7 days. The proposed approach is further validated by comparing its performance with commonly adopted machine learning techniques such as Random Forest, XGBoost, and Multilayer Perceptron, demonstrating its advantages in predicting player purchases.