Performance comparison.

Research on user churn prediction has been conducted across various domains for a long time. Among these, the gaming domain is characterized by its potential for diverse types of interactions between users. Due to this characteristic, many studies on churn prediction have considered the relationships between users and have primarily applied social network analysis. Recently, the use of Graph Neural Networks (GNNs) has been actively applied. However, existing studies utilizing GNNs have limitations as they use static graphs that do not effectively capture the dynamic nature of interactions that change over time. This study addresses these limitations by proposing a dynamic graph model for predicting user churn in games based on user interactions. Data are sourced from 10,000 users of ’Blade & Soul’ by NCSOFT. The proposed model effectively captures changes in user behavior over time and predicts user churn with a focus on interactions among users. Experimental results reveal that the proposed model achieves a higher F1 score compared with conventional algorithms and static graph models. Dynamic graphs more accurately reflect changes in user behavior compared with static graphs, particularly in domains with active interactions such as massively multiplayer online role-playing games. This work highlights the significance of user churn prediction in the gaming industry and demonstrates the effectiveness of the predictive models that use dynamic graphs.用户流失预测的相关研究长期以来已在诸多领域广泛开展。其中，游戏领域的显著特征是用户间可产生多样化的交互行为。鉴于这一特征，诸多流失预测研究均会考量用户间的关联关系，并主要采用社交网络分析方法。近年来，图神经网络（Graph Neural Networks, GNNs）的应用也愈发活跃。然而，现有采用GNN的研究存在局限性：其所用的静态图无法有效捕捉随时间变化的交互动态特性。本研究针对上述局限，提出一种基于用户交互行为的动态图模型，用于游戏场景下的用户流失预测。本研究的数据取自NCSoft开发的《剑灵》（Blade & Soul）的10000名用户。所提模型可有效捕捉用户行为随时间的变化，并基于用户间的交互关系实现用户流失预测。实验结果表明，相较于传统算法与静态图模型，所提模型的F1值更高。与静态图相比，动态图更能准确反映用户行为的变化，尤其在大型多人在线角色扮演游戏这类交互活跃的领域中。本研究凸显了游戏领域用户流失预测的重要性，并验证了采用动态图的预测模型的有效性。