Predição da evasão de jogadores em jogo grátis-para-jogar utilizando game analytics.

Abstract

The issue of player churn in free-to-play games poses a significant challenge in the electronic gaming industry. The growing popularity of these business models, where players can access the game for free, places a crucial emphasis on retaining these users to ensure the financial success and sustainability of the game. In this scenario, predictive analysis emerges as an essential tool to anticipate and understand the patterns of player churn. This study began with a systematic literature review in the field of predictive models in game analytics, aiming to answer the main research question: How are predictive models applied in game analytics? The research was conducted based on a protocol that defined the objectives, research questions, and inclusion and exclusion criteria. The main findings indicate that research on predictive models in game analytics has grown significantly since 2010, with a variety of machine learning techniques being applied. Furthermore, the most investigated prediction targets include the probability of winning, churn prediction, and player expertise. Regarding preprocessing techniques, several approaches were identified, such as Principal Component Analysis (PCA) and web scraping techniques. We focused our research on churn prediction, initially by defining churn and establishing cutoff dates, considering multiple time windows for classifying players as churned or recurrent. The analysis addressed the validity threats of the work, including churn definition issues, class imbalance, and the use of techniques like SMOTE to balance the data. Six machine learning models were evaluated, with an emphasis on metrics like accuracy, precision, recall, and AUC (Area Under the Curve). The 10-fold cross-validation technique was applied to validate the models, providing a more comprehensive view of their performance. The analysis of feature importance revealed which player characteristics were most relevant for churn prediction, although the interpretation of these features was highlighted as context-dependent. Ultimately, the work offered promising insights into the prediction of player churn in free-to-play games but emphasized the need for careful approaches and contextual considerations to mitigate validity threats and ensure the generalization of models to different datasets and time periods.