Customer Churn Prediction System: A Guided Web Work-flow for Local Training, Prediction, and Explainable Out-puts

Department of Computer Science and Engineering, CMR Technical Campus, Hyderabad, Telangana, India

Abstract

Customer churn—the event of a customer discontinuing a service—directly affects revenue and retention planning. This paper presents a Customer Churn Prediction System implemented as a guided web workflow that turns churn modeling into an interactive, repeatable pipeline. The plat-form supports dataset upload, preprocessing with configurable feature selection, model training, and interactive prediction through a step-based user interface. A FastAPI backend coordinates the machine learning stages and persists intermediate artifacts (uploaded files, preprocessing metadata, and trained models) for repeatable use. The predictive component uses an XGBoost classifier trained on structured customer attributes after standard tabular preprocessing (missing-value treatment and categorical encoding). At inference time, the system returns a churn probability and maps it to a discrete risk band using simple threshold rules to support operational decisions. The results interface surfaces key contributing factors via model feature importance and reports common evaluation metrics (accuracy, precision, recall, and F1score) for transparency. For interpretability, the backend can optionally invoke a locally hosted large LM through Ollama to convert the computed signals (probability, risk band, and top factors) into a concise explanation and recommended actions, while keeping data local. Overall, the system demonstrates how churn prediction can be delivered as an end-to-end application with reproducible workflow support and interpretability-focused reporting.