AI Insight
This study introduces a Multiplex Social Influence framework that predicts player churn in online games by analyzing both explicit friendship networks and implicit teammate relationships through network embedding techniques. The approach combines these multiplex social network embeddings with traditional behavioral data, achieving high prediction accuracy of 93% and an AUC of 0.98 on real-world gaming data. Ablation studies and robustness tests confirmed that incorporating multiple types of social relationships significantly outperforms models using single relationship types or behavioral data alone.
Why it matters
Accurate churn prediction enables gaming companies to identify at-risk players earlier and design targeted retention strategies, potentially reducing revenue loss and improving player engagement. The multiplex network approach could be adapted to other domains where users maintain multiple types of social connections, such as social media platforms, e-commerce, or subscription services.
Understand the Science
by Ming Guo, Yan Liu
Accurately predicting player churn is crucial for designing effective retention strategies and driving revenue in the online gaming industry. While existing predictive models achieve reasonable performance relying on individual demographic and behavioral data, they largely overlook the multiplex social influence arising from diverse, overlapping player relationships. To address this gap, we propose a novel Multiplex Social Influence (MSI) framework for player churn prediction. Specifically, we construct multiplex social networks encompassing both explicit friendship ties and implicit Player-versus-Environment (PvE) teammate relationships. We then apply a network embedding algorithm called MLNRL to efficiently extract multi-scale structural information from these distinct networks. These embeddings are concatenated to form a unified multiplex social influence representation and combined with players’ behavioral attributes for downstream classification. Extensive experiments on a real-world online game dataset validate the superiority of our proposed approach. Quantitatively, the MSI framework paired with an SVM (RBF kernel) classifier achieves outstanding prediction performance (accuracy: 0.93, F1-score: 0.93, AUC: 0.98. Furthermore, rigorous ablation studies, comprehensive robustness tests (covering network embedding algorithms, fusion strategies, and data imbalance issues), and parameter sensitivity analyses collectively confirm that integrating multiplex social influence yields robust synergistic effects, significantly outperforming baseline models built on single relationship types. These findings definitively confirm the effectiveness and necessity of incorporating multiplex social structures into data-driven churn prediction models for online game services.
Source: Online game player churn prediction based on multiplex social influence through network embedding