NobleBlocks

Research indexConference paper

Explainable Churn Analytics Using Machine Learning With Optimized Hyperparameter Tuning

Ramachandran, Kosaraju, Manideep and 1 other

26 Feb 2026

NobleIDark:/48914/w1/20261105/C7ECAAFCDOI10.1109/i5cps67958.2026.11452276

Paywalled

The full text is at the publisher

You can read the abstract here, or open it at the publisher.

Read at the publisher

Abstract

Telecom companies lose significant revenue each year when customers leave without warning, so our project focuses on building Telecom Churn Insights, an end-to-end system that converts raw telecom data into practical tools for customer retention. We start by thoroughly cleaning the IBM Telco churn dataset—removing identifiers, satisfaction-related fields, and any information recorded after churn to avoid leakage. After creating stratified train–validation–test splits, we train multiple classical machine-learning models, including Logistic Regression, Random Forest, XGBoost, CatBoost, and Naive Bayes, and fine-tune them using grid-search hyperparameter optimization. Our best model reaches an AUC of around 0.92.To interpret the predictions, we use SHAP for overall churn drivers—such as contract duration, payment method, and monthly charges—and LIME to explain individual customer predictions for retention teams. All these insights feed into a structured knowledge base that stores churn categories, reasons, and their frequencies. On top of this, a hybrid recommendation module produces action plans: fixed rule-based suggestions for known scenarios and context-aware recommendations generated by a Gemini LLM using the same knowledge base.

Discussion

Citations

More

Suggest a correction or report