DeepSeqCaus: A Deep Sequential Causal Inference Framework for User Churn Prediction and Optimal Retention Intervention Generation

Abstract

In digital platforms, understanding and mitigating user churn is crucial for sustaining long-term engagement and revenue. Traditional machine learning approaches often rely on correlation-based predictive models without explicitly accounting for causal relationships underlying user behavior. This study proposes DeepSeqCaus, a unified deep sequential causal inference framework that integrates sequence modeling and treatment effect estimation to enable accurate churn prediction and optimal retention intervention policy generation. DeepSeqCaus consists of a dual-branch architecture: (1) a Temporal Feature Encoder using gated convolution and bidirectional gated recurrent networks to extract multi-granular temporal representations from behavioral sequences; and (2) a Causal Effect Estimator based on counterfactual representation learning to estimate heterogeneous treatment effects (HTEs) for candidate interventions such as personalized notifications, discount offers, or content recommendations. Using large-scale user interaction logs from an online service, we conduct extensive experiments comparing DeepSeqCaus with conventional predictive models and causal inference baselines. The results showed that DeepSeqCaus outperformed the baseline model in all cases. The proposed framework provides actionable insights for targeted retention and demonstrates strong potential for deployment in intelligent customer management systems.