📱 Telecom Customer Churn Prediction

📌 Project Overview

Telecom operator "TeleDom" aims to reduce customer churn by offering personalized promocodes and special conditions to clients planning to leave. This project develops a machine learning model to predict the probability of contract termination.

🎯 Objective

Create a model that accurately predicts customer churn with:

• ROC-AUC ≥ 0.85

• High recall (prioritizing identifying potential churners over false positives)

📊 Dataset Description

Four CSV files containing customer information as of February 1, 2020:

1. contract_new.csv - contract details:

  - customerID, BeginDate, EndDate (target feature)

  - Billing and payment information (Type, PaperlessBilling, PaymentMethod)

  - Charges (MonthlyCharges, TotalCharges)

2. personal_new.csv - customer demographics:

  - Dependents, SeniorCitizen, Partner

3. internet_new.csv - internet services:

  - InternetService (DSL/Fiber optic)

  - Additional services (OnlineSecurity, OnlineBackup, etc.)

4. phone_new.csv - phone services:

  - MultipleLines

🔍 Methodology

Data Preparation

• Merged datasets using customerID

• Converted EndDate to binary target (1 = churned, 0 = active)

• Created new feature Months (contract duration)

• Handled missing values and transformed categorical features

• Removed highly correlated TotalCharges feature (MSE = 103.45 with MonthlyCharges * Months)

Modeling Approach

Tested multiple models with cross-validation and focal loss to address class imbalance (only ~26% churn rate):

• DecisionTreeClassifier (ROC-AUC: 0.808)

• RandomForestClassifier (ROC-AUC: 0.834)

• LGBMClassifier with focal loss (ROC-AUC: 0.899)

• LogisticRegression (ROC-AUC: 0.757)

• Neural Network (ROC-AUC: 0.871)

📈 Results

Best Model: LGBMClassifier with matched hyper-parameters.

Test Performance:

• ROC-AUC: 0.913

• Recall: 90% (correctly identifies 90% of customers planning to leave)

The model successfully identifies high-risk customers, enabling targeted retention strategies while minimizing customer loss.