Purpose This research aims to develop a deep learning methodology to predict customer churn in platform businesses while addressing the challenge of imbalanced data, using real-world data from a private chat platform to demonstrate significant improvements in churn prediction accuracy and model performance. Methods A comprehensive churn prediction methodology for predicting customer churn using CTGAN for data balancing and TabNet for model construction. This combined approach significantly enhances both the prediction accuracy and the interpretability of churn prediction models. Results The application of CTGAN and TabNet resulted in improved prediction accuracy and provided robust solutions for customer retention strategies. This research demonstrated the effectiveness of these advanced techniques in handling imbalanced data and enhancing model interpretability. The real-world case study involving data from a private chat platform underscored the practical value and impact of this methodology in real business scenarios. Conclusion This research presents an effective solution for churn prediction and valuable insights in platform businesses by utilizing deep learning methodology to manage imbalanced data and interpret model outcomes. The findings offer valuable insights for developing customer retention strategies. The approach provides actionable insights that can be applied to enhance customer retention and satisfaction strategies.