Customer churn prediction on MTN Nigeria dataset using TensorFlow, Pandas, and SHAP explainability
Customer Churn Prediction Model — MTN Nigeria
🔹 Overview
This project builds a machine learning prediction model to identify telecom customers at risk of churning. Using the MTN Nigeria customer dataset, the analysis explores key churn drivers and applies TensorFlow to predict churn probability, enabling targeted retention strategies.
🔹 Objectives
Analyze customer behavior to identify factors driving churn.
Build a baseline model (Logistic Regression) for comparison.
Develop a TensorFlow Neural Network to improve prediction accuracy.
Apply SHAP explainability to understand feature importance.
Deliver actionable business insights for churn reduction.
🔹 Dataset
Source: Kaggle – MTN Nigeria Customer Churn Dataset
(
kaggle.com)
Size: ~974 customers
Features: Age, Gender, Device, Subscription Plan, Tenure, Satisfaction Rate, Revenue, Data Usage, etc.
Target: Customer Churn Status (1 = churned, 0 = stayed)
🔹 Methodology
1) Data Cleaning & Preprocessing
Standardized target labels (0/1).
Dropped IDs and free-text columns.
Handled missing values with imputation.
One-hot encoded categorical variables.
2) Exploratory Data Analysis (EDA)
Distribution of churn (29% churned, 71% retained).
Boxplots and bar charts showing churn differences across revenue, tenure, satisfaction, and subscription plans.
Correlation heatmaps for numeric features.
3) Modeling
Baseline: Logistic Regression (ROC-AUC ~0.75).
Neural Network: TensorFlow MLP with early stopping (ROC-AUC ~0.86).
4) Model Evaluation
Accuracy, ROC-AUC, Confusion Matrix.
Precision-Recall curve for threshold tuning.
5) Explainability
SHAP analysis confirmed key drivers: short tenure, low satisfaction rate, subscription type, revenue/usage levels.
🔹 Results
Baseline (Logistic Regression): ROC-AUC = ~0.75
TensorFlow Model: ROC-AUC = ~0.86
Key Insight: Short-tenure, low-satisfaction, and prepaid customers have the highest churn risk …