This project predicts subscriber churn for a music streaming service (KKBox) and turns each prediction into a plain-language explanation for a retention team to act on.

Starting from 23 million raw transactions and usage records, we built 25 monthly cohorts spanning two years to create a time-agnostic model. We engineered behavioral features capturing user velocity.

We compared five models using a time-based train/validation/test split and PR-AUC as the metric. We also evaluated a weighted ensemble of these five models, weighted based on the time-series CV. Since the ensemble performance only offers a marginal gain in PR-AUC on the validation set and loses the interpretability from XGBoost, we selected XGBoost as the final model. XGBoost achieved a 0.542 PR-AUC on the test data.

We added a final layer using SHAP to explain a given user's prediction and an LLM to translate into a retention report. The structured SHAP output acts as a factual guardrail against hallucinated reasoning.