Modern Economics & Management Forum

Based on the negative comment text of mobile applications, this research proposes a framework integrating Latent Dirichlet Allocation (LDA) with duration modeling to predict not only when users churn but also the underlying reasons for their departure. The study aims to address a gap in existing literature, which often focuses solely on churn probability or timing, by incorporating thematic analysis of user-generated content from app stores. Methodologically, it employs the Duration Model to estimate the time until churn and extends it into a Competitive Risk Model that accounts for multiple churn reasons (categorized as controllable, uncontrollable, and unknown risks). The LDA algorithm is utilized to extract latent topics from negative reviews, transforming unstructured text into interpretable variables for predictive modeling. These variables, alongside duration data, are integrated into the model to enhance prediction accuracy. Evaluation will involve metrics such as perplexity for LDA performance and log-likelihood, AIC, and BIC for model comparison. The research expects to contribute a novel, text-enhanced approach to churn prediction, offering practical insights for internet companies to better understand churn drivers, improve user experience, and design targeted retention strategies, despite potential limitations in LDA's performance on short text.