Advanced customer churn prediction for a music streaming digital marketing service using attention graph-based deep learning approach.

Customer churn poses a persistent threat to the sustainability of music streaming platforms, where user disengagement often occurs unpredictably. Most existing churn prediction methods fail to integrate relational dependencies among users or address the limitations of static, tabular models in imbalanced datasets. In this research, we propose a Hybrid Graph Attention Network (Hybrid GAT + MLP (Multi Layered Perceptron)) that efficiently combines graph-based learning with deep tabular feature modelling, allowing the capture of both individual behaviour and peer-influenced churn patterns. The use of a synthetic similarity graph enables the model to learn from neighbourhood context, while a weighted loss function ensures robustness against class imbalance without altering data distribution. The model was evaluated on the KKBox dataset, achieving an accuracy of 95.8% with an AUC score of 0.9626, indicating high discriminative power. Additional metrics, such as F1-score, confusion matrix, and ROC analysis, ensure the system’s ability to generalize across both churn and non-churn classes. These results validate the applicability of our approach to large-scale, real-world scenarios involving churn detection. All code and the curated dataset for customer churn prediction for music steaming is available at the https://github.com/haiyanchenas/Advanced-Customer-Churn-Prediction-for-a-Music-Streaming-Digital-Marketing-Service/tree/main enabling reproducibility of our findings.